Blood vessel three-dimensional reconstruction method and system based on digital subtraction angiography

Abstract

The invention provides a blood vessel three-dimensional reconstruction method and system based on digital subtraction angiography, and relates to the technical field of medical image processing and three-dimensional reconstruction. The method comprises the following steps: performing frame-by-frame registration, subtraction processing and uniform sampling on a collected two-dimensional digital subtraction angiography DSA image sequence to obtain a frame sparse view; constructing and training a vessel three-dimensional reconstruction network based on the sparse view angle dynamic DSA, inputting the frame sparse view into the trained vessel three-dimensional reconstruction network based on the sparse view angle dynamic DSA, and generating geometric parameters and center attenuation values of all Gaussian kernels; calculating three-dimensional blood vessel volumes under all timestamps and performing voxelization averaging on the three-dimensional blood vessel volumes under all timestamps on the basis of the geometric parameters and the center attenuation values of all Gaussian kernels, and generating a static three-dimensional blood vessel model; and meanwhile, all Gaussian kernels are rendered to generate a synthesized DSA image under a new view angle. The invention aims to improve the integrity and the time sequence consistency of the micro blood vessel reconstruction under the sparse view angle.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and three-dimensional reconstruction technology, and in particular to a method for three-dimensional reconstruction of blood vessels based on digital subtraction angiography (DSA). Background Technology

[0002] Digital subtraction angiography (DSA) is currently the standard for clinical diagnosis of vascular diseases such as arteriovenous malformations, arteriovenous fistulas, and intracranial aneurysms. Its imaging principle involves performing rotating cone-beam X-ray scans before and after contrast agent injection (i.e., masking and filling scans), and then logarithmically subtracting the two sets of images to obtain a two-dimensional DSA image that eliminates background structures such as bone and soft tissue, clearly highlighting the dynamically filling vascular lumen with blood flow. However, the overlapping of vascular structures in two-dimensional DSA images severely limits the accurate assessment of complex three-dimensional vascular anatomy by physicians. Therefore, reconstructing three-dimensional vascular structures from DSA image sequences is crucial for precision medical diagnosis and surgical planning.

[0003] Currently, commercial DSA systems typically require the acquisition of hundreds of projection images to achieve high-quality 3D reconstruction, relying on cone-beam computed tomography (CBCT) reconstruction algorithms such as Feldkamp-Davis-Kress (FDK). This intensive acquisition mode results in significant radiation doses for both patients and healthcare professionals. To mitigate radiation risks, sparse-view DSA 3D reconstruction (using only a few dozen images) has become an important research direction, but it faces two major challenges: First, the dynamic imaging challenge. DSA image sequences capture the instantaneous state of contrast agent flow in blood vessels, with each frame corresponding to a different degree of vascular filling, contradicting the assumption of a static scene in traditional CT. Second, the challenge of reconstructive pathology. In cases of severely insufficient projection data, static reconstruction algorithms such as FDK produce severe fringe artifacts and noise, leading to a sharp decline in reconstruction quality.

[0004] In recent years, implicit neural representation methods such as Neural Radiance Fields (NeRF) have provided new ideas for dynamic sparse view reconstruction. These methods encode complex scenes into a continuous implicit function through small neural networks, enabling the synthesis of high-fidelity, high-resolution new views from a finite or even sparse number of input images. The paper "3D vessel reconstruction from sparse-viewdynamic DSA images via vessel probability guided attenuation learning[J].arXiv preprint arXiv:2405.10705, 2024." proposes that attenuation learning guided by vessel probability addresses the aforementioned challenges to some extent and achieves better reconstruction results. However, these methods use multilayer perceptrons (MLPs) to model the entire imaging space, requiring redundant computation on a large number of empty background regions, resulting in an extremely time-consuming training process. A single case often requires several hours, making it difficult to meet the real-time efficiency requirements of clinical practice.

[0005] Meanwhile, 3D Gaussian Splatting (3DGS) technology has attracted widespread attention in the field of computer vision due to its efficiency in explicit scene representation and differential rasterization rendering. Compared with NeRF volumetric rendering, kernel-based explicit representation can naturally avoid empty background regions, concentrate computational resources on the target structure, and rasterization rendering is much faster than volumetric rendering. This makes it very suitable for fast and clear 3D reconstruction of vascular networks with sparse morphology, fine structure, and complex background. Several studies have attempted to introduce Gaussian splashing into the field of X-ray imaging. For example, Zha R, LinT J, Cai Y, et al. R 3D-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction[J]. arXiv preprint arXiv:2405.20693, 2024. first applied it to static CT reconstruction, proposing the concept of "radiative Gaussian splashing" and developing X-ray rasterization and voxelization methods. Zhang S, Zhao H, Zhou Z, et al. Togs: Gaussian splatting with temporalopacity offset for real-time 4d dsa rendering[J]. IEEE Journal of Biomedical and Health Informatics, 2025. applied it to the dynamic synthesis of DSA images. However, few studies have directly applied Gaussian splashing technology to three-dimensional vascular reconstruction from sparse dynamic DSA images.

[0006] However, current methods for 3D vascular reconstruction using sparse-view dynamic DSA struggle to effectively preserve the geometric structure information of small blood vessels during feature extraction. Furthermore, each Gaussian kernel learns temporal encoding independently, neglecting the spatial prior that neighboring kernels within the same vascular segment should share similar hemodynamic features. This fragmented modeling approach not only results in significant computational redundancy, limiting further efficiency improvements, but also fails to explicitly utilize the temporal consistency of vascular segments, leading to potentially uneven and inconsistent reconstructed dynamic vessels. It also hinders the effective learning of robust blood flow patterns from training data, impacting generalization ability and final reconstruction accuracy under extremely sparse perspectives. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention uses a two-dimensional DSA image sequence acquired by rotation to automatically reconstruct and characterize the three-dimensional spatial morphology of the vascular structure and its decay characteristics over time during the entire contrast agent flow cycle, thereby establishing a complete and accurate static three-dimensional vascular model. This invention proposes a three-dimensional vascular reconstruction method based on digital subtraction angiography, aiming to improve the integrity and temporal consistency of microvascular reconstruction under sparse perspective.

[0008] On the one hand, this invention proposes a method for three-dimensional reconstruction of blood vessels based on digital subtraction angiography, which includes the following process:

[0009] Acquire a sequence of two-dimensional digital subtraction angiography (DSA) images of vascular regions obtained through rotational acquisition;

[0010] The VoxelMorph non-rigid image registration method based on deep learning is used to register the two-dimensional digital subtraction angiography (DSA) image sequence frame by frame, and the registered image sequence is then subjected to subtraction processing to generate the subtracted vascular sequence.

[0011] Uniform sampling was performed on the subtraction angiography sequence to obtain... Frame sparse view;

[0012] A 3D vascular reconstruction network based on sparse-view dynamic DSA was constructed and trained to obtain a well-trained 3D vascular reconstruction network based on sparse-view dynamic DSA.

[0013] Will The frame sparse view is input into the trained vascular 3D reconstruction network based on sparse view dynamic DSA, generating the geometric parameters of each Gaussian kernel and the center decay value at all timestamps;

[0014] Based on the geometric parameters of each Gaussian kernel and the center decay value at all timestamps, the three-dimensional blood vessel volume is calculated at all timestamps.

[0015] By performing voxelized averaging on the three-dimensional blood vessel volumes at all timestamps, a static 3D blood vessel volume is obtained, which in turn generates a static three-dimensional blood vessel model.

[0016] Based on the geometric parameters of all Gaussian kernels and the center attenuation value at all timestamps, a differentiable X-ray rasterization model is used to render all Gaussian kernels, generating a synthetic DSA image from a new perspective.

[0017] Furthermore, the specific content of acquiring the two-dimensional digital subtraction angiography (DSA) image sequence of the vascular region obtained through rotation acquisition is as follows:

[0018] Without injecting contrast agent, a series of continuous two-dimensional background images are acquired through the first rotation along a preset rotation trajectory, which is denoted as the mask sequence.

[0019] After the contrast agent is injected into the blood vessel, a second rotational acquisition is performed along the same rotational trajectory as in the first rotational acquisition, to obtain a set of continuous two-dimensional dynamic contrast sequences, denoted as the filling sequence;

[0020] The acquisition parameters, viewpoint, and frame rate of the second rotation acquisition are kept consistent with those of the first rotation acquisition to ensure that the first [part of the] filling sequence is complete. The frame image and the first in the mask sequence The frame images satisfy the requirement of consistent inter-frame correspondence;

[0021] The mask sequence and filling sequence are used as two-dimensional digital subtraction angiography (DSA) image sequences of the vascular region.

[0022] Furthermore, the VoxelMorph non-rigid image registration method based on deep learning is used to register the two-dimensional digital subtraction angiography (DSA) image sequence frame by frame, and the registered image sequence is then subjected to subtraction processing to generate the subtracted vascular sequence. The specific content of this subtracted vascular sequence is as follows:

[0023] Obtain several pairs of masked frames and filling frames, stitch each pair of masked frames and filling frames together in the channel dimension, and use the stitched two-channel tensor to construct a training set.

[0024] For any dual-channel tensor in the training set, the dual-channel tensor is input into a convolutional neural network, and a two-dimensional displacement vector field (DVF) is output. Based on the two-dimensional displacement vector field (DVF), a spatial transformation network is used to register the filling frames corresponding to the dual-channel tensor to obtain the registered filling image.

[0025] Construct the joint loss of the convolutional neural network, and optimize the network parameters of the convolutional neural network by minimizing the joint loss;

[0026] For any timestamp The corresponding mask frames and filling frames are extracted from the two-dimensional digital subtraction angiography (DSA) image sequence. and extract the mask frame and full frames By concatenating the components along the channel dimension, a two-channel tensor is obtained.

[0027] The dual-channel tensor is input into the optimized convolutional neural network, which outputs a two-dimensional displacement vector field (DVF).

[0028] Based on the two-dimensional displacement vector field (DVF), a spatial transformation network is used to fill the frames. Registration is performed to obtain a registered full image. ;

[0029] Calculate the registered filling images separately. and mask frame The logarithmic domain pixel values, and the registered full image With mask frame Subtracting pixel by pixel yields the timestamp. Image of blood vessels after lower subtraction;

[0030] Based on the subtracted vascular images at all timestamps, generate the subtracted vascular sequence.

[0031] Furthermore, the vascular three-dimensional reconstruction network based on sparse-view dynamic DSA includes:

[0032] The preliminary 3D reconstruction module is used to perform FDK algorithm on the input. The sparse view of the frame is used for 3D reconstruction to generate a 3D blood vessel volume model composed of several individual pixels.

[0033] The 4D Gaussian initialization module is used to perform non-empty voxel sampling on the three-dimensional blood vessel volume model, and construct an initial Gaussian kernel set based on the voxel points obtained from the sampling, while initializing the geometric parameters and dynamic decay parameters of each Gaussian kernel in the set.

[0034] The cluster-shared real-time coding module is used to generate a fused temporal feature vector characterizing the dynamic characteristics of a blood vessel segment based on the geometric parameters, dynamic decay parameters, and timestamps of the Gaussian kernel, through parallel cluster-level shared coding pathways and kernel-level residual compensation pathways.

[0035] The kernel-level MLP module with residuals is used to perform feature transformation on the fused temporal feature vector characterizing the dynamic properties of vascular segments, generating the center decay value of each Gaussian kernel at a specific time.

[0036] Furthermore, the geometric parameters of the Gaussian kernel include: spatial position. Rotation parameters and scale parameters Dynamic attenuation parameters, including: cluster coding matrix ;

[0037] The specific contents of the 4D Gaussian initialization module are as follows:

[0038] Iterate through all voxel points in the three-dimensional blood vessel volume model, and match each voxel point with a preset voxel threshold. Compare and retain all voxel values ​​greater than 1. voxel points;

[0039] Each retained voxel point is used as a Gaussian kernel to construct an initial set of Gaussian kernels;

[0040] For each Gaussian kernel in the initial Gaussian kernel set, the spatial position of the voxel point corresponding to the Gaussian kernel is taken as the initial spatial position of the Gaussian kernel.

[0041] The rotation parameters of the Gaussian kernel are initialized to a unit quaternion. This indicates that the Gaussian kernel did not rotate in the initial state;

[0042] The Gaussian kernel that is closest in the initial point cloud Calculate the Gaussian kernel to the nearest neighbor point. The average distance between the neighboring points is used to calculate the initial scale parameters of the Gaussian kernel.

[0043] The Kaiming initialization strategy is adopted to initialize the weight parameters of the cluster-shared real-time coding CTEM module and the kernel-level MLP module with residuals;

[0044] Clustering algorithms are used to divide all Gaussian kernels in the initial Gaussian kernel set into several spatial clusters, each representing a local blood vessel segment. Cluster labels are then assigned to each Gaussian kernel based on the partitioning results. ;

[0045] Based on the cluster label A cluster encoding matrix is ​​initialized for each partitioned spatial cluster. .

[0046] Furthermore, the specific content of the cluster-shared real-time encoding module is as follows:

[0047] For any Gaussian kernel in a spatial cluster The spatial position of the Gaussian kernel Cluster tags and timestamp As input data, the input data is encoded through parallel cluster-level shared coding pathways and kernel-level residual compensation pathways to generate fusion temporal feature vectors characterizing the dynamic properties of vascular segments;

[0048] The cluster-level shared coding path is as follows: for Gaussian kernels To determine the spatial cluster to which it belongs, obtain the cluster encoding matrix of that spatial cluster. Based on the time basis function, according to the timestamp Generate weight vector The weight vector mentioned above The length of the cluster coding matrix number of rows;

[0049] Based on the weight vector Each element in the cluster encoding matrix The corresponding rows in the vector are scaled, and then all the scaled row vectors are added together to obtain the cluster-level shared coding vector of the spatial cluster.

[0050] The kernel-level residual compensation path is as follows: using a 3D hash encoder to perform Gaussian kernel... spatial location Encode and generate Gaussian kernel Spatial eigenvectors; Gaussian kernel timestamp Perform sine coding to generate a Gaussian kernel. Time feature vector;

[0051] Gaussian kernel The spatial and temporal feature vectors are concatenated and input into the residual prediction MLP for learning and prediction, generating a Gaussian kernel. Kernel-level residual compensation vector ;

[0052] Cluster-level shared coding vector and kernel-level residual compensation vector The data are then concatenated to generate a fused temporal feature vector that characterizes the dynamic properties of the vascular segment.

[0053] Furthermore, the kernel-level MLP module with residuals is a lightweight multilayer perceptron containing residual connections, comprising two fully connected units in series; each fully connected unit includes a fully connected layer (FC) and a ReLU activation function in series; the outputs of the two fully connected units are fused through residual connections and then input into a third fully connected layer (FC3) for dimension mapping to obtain the time stamp of each Gaussian kernel. center attenuation value .

[0054] Furthermore, the training process of the vascular 3D reconstruction network based on sparse viewpoint dynamic DSA is as follows:

[0055] Build includes The training samples of the sparse view are input into the vascular 3D reconstruction network based on sparse view dynamic DSA for iterative training. In the first iteration, the training samples are input into the vascular 3D reconstruction network based on sparse view dynamic DSA to obtain an initial Gaussian kernel set. From the second iteration onwards, the Gaussian kernel set generated after the previous iteration is used as the optimization object for each iteration. The Gaussian kernel set includes: spatial location... Rotation parameters Scale parameters Cluster coding matrix And the network weight parameters of the cluster-shared real-time coding module and the KR-MLP module;

[0056] For each iteration in the training process, the following procedure is performed:

[0057] Based on the current Gaussian kernel set, the center attenuation values ​​of all Gaussian kernels are predicted using a cluster-shared real-time encoding module and a KR-MLP module. Then, a differentiable X-ray rasterization model is used to render the current Gaussian kernel set into a synthetic DSA image. ;

[0058] Determine if the current iteration round has reached the preset maximum number of iterations. If it has, obtain the trained 3D blood vessel reconstruction network based on sparse viewpoint dynamic DSA using the optimized Gaussian kernel set for the current iteration round. If it has not reached the maximum number of iterations, calculate the synthesized DSA image. Compared with the real DSA image under the current sparse view The loss value between the two is used to update the Gaussian kernel set in the current iteration through the backpropagation algorithm;

[0059] After completing a fixed iteration interval, pruning, cloning, or splitting operations are performed on the Gaussian kernel set under the current iteration round to complete the adaptive adjustment of the Gaussian kernel set under the current iteration round.

[0060] Furthermore, the specific details of calculating the three-dimensional blood vessel volume at all time points based on the geometric parameters of each Gaussian kernel and the center decay value at all time points are as follows:

[0061] When using the FDK algorithm on the input During the 3D reconstruction of sparse frame views, the region of interest (ROI) is determined.

[0062] For any non-empty voxel sampling point in the region of interest (ROI) Based on non-empty voxel sampling points Corresponding Gaussian kernel Geometric parameters, calculate non-empty voxel sampling points Spatial Gaussian distribution function This function value is used to represent non-empty voxel sampling points. Weights in the region of interest (ROI);

[0063] For any timestamp According to Gaussian kernel In timestamp center attenuation value Calculate timestamps lower non-empty voxel center point The overall prime value;

[0064] timestamp Perform voxel integration on the total voxel values ​​of all non-empty voxel centers to obtain the voxel values ​​at the timestamp. Three-dimensional blood vessel volume .

[0065] On the other hand, this invention proposes a three-dimensional vascular reconstruction system based on digital subtraction angiography, the system comprising:

[0066] The image sequence acquisition module is used to acquire two-dimensional digital subtraction angiography (DSA) image sequences of vascular regions obtained through rotational acquisition.

[0067] The registration and subtraction module uses the VoxelMorph non-rigid image registration method based on deep learning to register the two-dimensional digital subtraction angiography DSA image sequence frame by frame, and then performs subtraction processing on the registered image sequence to generate the subtracted vascular sequence.

[0068] The uniform sampling module is used to uniformly sample the subtracted vascular sequence to obtain... Frame sparse view;

[0069] The 3D reconstruction module is used to call the trained vascular 3D reconstruction network based on sparse viewpoint dynamic DSA, and to... The frame sparse view is input into the trained vascular 3D reconstruction network based on sparse view dynamic DSA, generating the geometric parameters of each Gaussian kernel and the center decay value at all timestamps;

[0070] The temporal volume generation module is used to calculate the three-dimensional blood vessel volume at all time points based on the geometric parameters of each Gaussian kernel and the center decay value at all time points.

[0071] The static model generation module is used to obtain the static 3D blood vessel volume by voxel averaging the 3D blood vessel volume at all timestamps, and then generate a static 3D blood vessel model.

[0072] The rendering and compositing module is used to render all Gaussian kernels using a differentiable X-ray rasterization model based on the geometric parameters of all Gaussian kernels and the center attenuation value at all timestamps, generating a synthetic DSA image from a new perspective.

[0073] The beneficial effects of adopting the above technical solution are as follows:

[0074] To address the problems of computational redundancy in temporal modeling, poor consistency of blood flow dynamics within vessel segments, and insufficient integrity in the reconstruction of microvessels in existing sparse-view reconstruction methods, this invention utilizes a cluster-level coding framework to uniformly model hemodynamic features at the vessel segment level during training. Since there is a strong spatiotemporal correlation between the blood flow dynamics of neighboring Gaussian kernels and their respective vessel segments, this invention designs a synergistic mechanism of cluster-level shared coding and kernel-level residual compensation to learn the temporal patterns of blood flow filling from both the overall and local fine-tuning dimensions of the vessel segment. This mechanism clusters Gaussian kernels according to their spatial location and vascular structure, learning a unified basic pattern of blood flow dynamics for each vessel segment, while preserving personalized temporal fine-tuning capabilities for each kernel within the cluster. On the one hand, it can explicitly utilize the spatiotemporal consistency prior of vessel segments to uniformly model hemodynamic features at the cluster level, significantly improving reconstruction efficiency and temporal consistency. On the other hand, the kernel-level residual compensation mechanism preserves the temporal changes of local vascular details, ensuring the precision of microvessel branch reconstruction while reducing computational complexity, providing an efficient and accurate vascular dynamic imaging solution for clinical use. Therefore, compared with previous methods, the method of the present invention has higher integrity and temporal consistency in the reconstruction of microvessels from a sparse perspective, while achieving several times the training acceleration.

[0075] This invention provides a 3D vascular reconstruction method based on sparse-view dynamic DSA. This method utilizes a collaborative mechanism of cluster-shared real-time encoding and kernel-level residual compensation, an adaptive clustering strategy based on spatial location and vascular centerline, and a deep fusion of cluster-level temporal features and kernel-level geometric features to achieve unified modeling of hemodynamics at the vascular segment level and accurate preservation of local details. This significantly improves the integrity and temporal consistency of microvascular reconstruction, increases training efficiency by several times, and greatly reduces the requirements for computing hardware, providing a feasible technical path for clinical intraoperative 3D vascular imaging. Attached Figure Description

[0076] Figure 1 This is a flowchart of a three-dimensional vascular reconstruction method based on digital subtraction angiography in this embodiment;

[0077] Figure 2 This is a schematic diagram of the process for three-dimensional reconstruction of blood vessels based on digital subtraction angiography in this embodiment;

[0078] Figure 3 This is a schematic diagram of the dynamic DSA image sequence for registration and subtraction processing in this embodiment; where (a) is a view of the filling sequence; (b) is a view of the masking sequence; and (c) is a view of the subtracted sequence.

[0079] Figure 4 This is a structural diagram of the vascular 3D reconstruction network based on sparse view dynamic DSA in this embodiment;

[0080] Figure 5 This is a schematic diagram of the cluster-shared real-time encoding module in this embodiment;

[0081] Figure 6 This is a schematic diagram of the KR-MLP module in this embodiment;

[0082] Figure 7 This is a schematic diagram of the new view synthesis and 3D reconstruction results in this embodiment; where (a) is a schematic diagram of the new view synthesis result; and (b) is a schematic diagram of the 3D reconstruction result.

[0083] Figure 8 This is a structural diagram of the three-dimensional vascular reconstruction system based on digital subtraction angiography in this embodiment. Detailed Implementation

[0084] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0085] Example 1:

[0086] This embodiment presents a method for three-dimensional vascular reconstruction based on digital subtraction angiography, such as... Figure 1 and Figure 2 As shown, the method includes the following steps:

[0087] Acquire a sequence of two-dimensional digital subtraction angiography (DSA) images of vascular regions obtained through rotational acquisition;

[0088] The specific content of obtaining the two-dimensional digital subtraction angiography (DSA) image sequence of the vascular region obtained through rotation acquisition is as follows:

[0089] Without injecting contrast agent, a series of continuous two-dimensional background images are acquired through the first rotation along a preset rotation trajectory, denoted as the mask sequence.

[0090] After the contrast agent is injected into the blood vessel, a second rotational acquisition is performed along the same rotational trajectory as in the first rotational acquisition, to obtain a set of continuous two-dimensional dynamic contrast sequences, denoted as the filling sequence.

[0091] In this embodiment, based on rotational acquisition, a C-arm is used to obtain a two-dimensional digital subtraction angiography (DSA) image sequence, which is then encapsulated in NIfTI (Neuroimaging Informatics Technology Initiative) format with the .nii extension to store time-series image pairs of the two-dimensional DSA mask run and fill run. The mask run refers to a sequence of two-dimensional background images continuously acquired at a fixed angle or along a preset rotation trajectory without contrast agent injection. The fill run refers to a two-dimensional dynamic angiography sequence acquired after contrast agent is injected into the blood vessel through a catheter, using the same acquisition parameters, viewpoint, and frame rate as the mask run. Both are directly acquireable data, and it is necessary to ensure consistent frame correspondence between the two acquisition processes to form frame-by-frame matched image pairs for subsequent subtraction processing.

[0092] The acquisition parameters, viewpoint, and frame rate of the second rotation acquisition are kept consistent with those of the first rotation acquisition to ensure that the first [part of the] filling sequence is complete. The frame image and the first in the mask sequence The frame images satisfy the requirement of consistent correspondence between frames.

[0093] In this embodiment, the acquisition parameters include, but are not limited to, exposure parameters, detector parameters, and device geometry parameters. Exposure parameters, such as tube voltage (kVp), tube current (mA), and exposure time (s) that control the energy, intensity, and dose of X-rays, must be consistent to ensure that the attenuation characteristics of background tissue (e.g., bone) remain the same in corresponding frames of two acquisitions. Detector parameters, such as image matrix size and pixel dimensions, define the spatial resolution and format of the acquired image. Device geometry parameters, such as focal distance to image receiver (FDD) and the center position of the C-arm, define the spatial relationship between the X-ray source, patient, and detector, ensuring that the image magnification, distortion, and projection geometry remain consistent.

[0094] In the DSA process, the rotation trajectory refers to the spatial path and angle sequence that the C-arm travels throughout the entire scanning cycle, while the viewing angle refers to the instantaneous projection angle corresponding to any frame of the image on the rotation trajectory. The consistent viewing angles of the two acquisitions mean that the second acquisition must completely or almost completely repeat the entire angle sequence of the first acquisition to ensure that the difference in acquisition angle between the frames corresponding to the two acquisitions does not exceed 0.5 degrees.

[0095] Maintaining a consistent frame rate is to ensure that the first frame in the filling sequence... The frame image and the first in the mask sequence The frame images satisfy the requirement of consistent correspondence between frames.

[0096] The mask sequence and filling sequence are used as two-dimensional digital subtraction angiography (DSA) image sequences of the vascular region.

[0097] The VoxelMorph non-rigid image registration method based on deep learning is used to register the two-dimensional digital subtraction angiography (DSA) image sequence frame by frame, and the registered image sequence is then subjected to subtraction processing to generate the subtracted vascular sequence.

[0098] In this embodiment, the VoxelMorph non-rigid image registration method based on deep learning is used to estimate and align the mask sequence and the filling sequence to obtain the registered image sequence.

[0099] The VoxelMorph non-rigid image registration method based on deep learning is used to register the two-dimensional digital subtraction angiography (DSA) image sequence frame by frame, and then the registered image sequence is subjected to subtraction processing to generate the subtracted vascular sequence.

[0100] Obtain several pairs of masked frames and filling frames, stitch each pair of masked frames and filling frames together along the channel dimension, and use the stitched dual-channel tensor to construct a training set.

[0101] For any dual-channel tensor in the training set, the dual-channel tensor is input into a convolutional neural network, which outputs a two-dimensional displacement vector field (DVF). Based on the DVF, a spatial transformation network is used to register the filling frames corresponding to the dual-channel tensor to obtain the registered filling image.

[0102] Construct the joint loss of the convolutional neural network and optimize the network parameters of the convolutional neural network by minimizing the joint loss.

[0103] In this embodiment, the CNN minimizes a joint loss function. Optimization is performed to achieve DVF estimation. The joint loss of the convolutional neural network is:

[0104] (1)

[0105] in Indicates joint loss; The similarity loss is represented by Local Normalized Cross-Correlation (NCC) loss, which is used to measure the similarity loss of the registered images. With mask Pixel similarity between them; These are the weighting coefficients; The regularization loss is represented by a smoothness loss based on the sum of squared spatial gradients of the DVF, which penalizes drastic changes in the DVF and ensures the continuity and physical rationality of the deformation field.

[0106] For any timestamp Extract the corresponding mask frame from the two-dimensional digital subtraction angiography (DSA) image sequence. and full frames and extract the mask frame and full frames By splicing along the channel dimension, a two-channel tensor is obtained.

[0107] The dual-channel tensor is input into the optimized convolutional neural network, which outputs a two-dimensional displacement vector field (DVF).

[0108] In this embodiment, the convolutional neural network (CNN) adopts an encoder-decoder structure similar to U-Net to predict the two-dimensional displacement vector field (DVF).

[0109] Based on the two-dimensional displacement vector field (DVF), a spatial transformation network is used to fill the frames. Registration is performed to obtain a registered full image. .

[0110] In this embodiment, the predicted DVF is applied to the filling frame using a Spatial Transformer Network (STN). This yields a registered, full-fill image. .

[0111] Calculate the registered filling images separately. and mask frame The logarithmic domain pixel values, and the registered full image With mask frame Subtracting pixel by pixel yields the timestamp. Image of blood vessels after lower subtraction;

[0112] Based on the subtracted vascular images at all timestamps, generate the subtracted vascular sequence.

[0113] In this embodiment, as Figure 3As shown, the input mask run and fill run sequences are registered and subtracted frame by frame to obtain a DSA image sequence of blood vessels. Specifically, the registration stage uses the existing deep learning-based VoxelMorph image registration method to estimate and align the mask run and fill run sequences using non-rigid deformation estimation, thereby minimizing spatial deviations caused by patient micromovements, respiration, and C-arm rotation. After registration, logarithmic subtraction is performed on each pair of corresponding frames. This involves calculating the logarithmic pixel values ​​of the mask and fill run images and then subtracting them pixel by pixel to suppress background tissue and highlight the attenuation changes of the contrast agent within the blood vessel lumen, ultimately generating the subtracted blood vessel sequence.

[0114] Uniform sampling was performed on the subtraction angiography sequence to obtain... Frame sparse view.

[0115] The uniform sampling method is as follows: the subtracted vascular sequence is sampled at equal time intervals or equal angular intervals to obtain... Frame sparse view.

[0116] In this embodiment, samples are uniformly taken from the subtracted vascular sequence at equal time intervals or equal angular intervals. Frames constitute the sparse view dataset used for model training, where It is a smaller subset of the total number of frames in the original sequence. For example, N usually does not exceed 30% to 40% of the total number of frames.

[0117] A 3D vascular reconstruction network based on sparse-view dynamic DSA was constructed and trained to obtain a well-trained 3D vascular reconstruction network based on sparse-view dynamic DSA.

[0118] In this embodiment, as Figure 4 As shown, a 3D vascular reconstruction network structure based on sparse view dynamic DSA is constructed based on the 3D Gaussian splash framework. This structure innovatively introduces a cluster-wise temporal encoding module (CTEM) and a kernel-level MLP with residuals (KR-MLP), which decouples the complex 3D vascular reconstruction task into two sub-tasks: static structural modeling and dynamic temporal modeling.

[0119] The vascular 3D reconstruction network based on sparse-view dynamic DSA includes:

[0120] The preliminary 3D reconstruction module is used to perform FDK algorithm on the input. The sparse view is used to perform 3D reconstruction, generating a 3D blood vessel volume model composed of several individual pixels.

[0121] In this embodiment, the network is equipped with registered subtraction and sparse view sampling ( The system takes a sequence of dynamic DSA blood vessel images (frames) as input. Then, the existing FDK algorithm is used to perform preliminary 3D reconstruction of the sparse view, generating a coarse blood vessel volume as the basic geometric prior. Subsequently, a 4D radiative Gaussian model is initialized within this basic volume.

[0122] The 4D Gaussian initialization module is used to perform non-empty voxel sampling on the three-dimensional blood vessel volume model, and construct an initial Gaussian kernel set based on the voxel points obtained from the sampling, while initializing the geometric parameters and dynamic decay parameters of each Gaussian kernel in the set.

[0123] The geometric parameters of the Gaussian kernel include: spatial position Rotation parameters and scale parameters The dynamic attenuation parameter includes: cluster coding matrix. .

[0124] The specific contents of the 4D Gaussian initialization module are as follows:

[0125] Iterate through all voxel points in the three-dimensional blood vessel volume model, and match each voxel point with a preset voxel threshold. Compare and retain all voxel values ​​greater than 1. voxel points.

[0126] Each retained voxel point is used as a Gaussian kernel to construct an initial set of Gaussian kernels.

[0127] In this embodiment, a preset voxel threshold is set. Iterate through all voxels in the 3D blood vessel volume model and compare their voxel values ​​with a threshold. Compare and retain all voxel values ​​greater than 100%. The voxel points. These preserved voxel points constitute the initial set of Gaussian kernels, and their spatial positions will be used as the initial spatial positions of each Gaussian kernel. This operation concentrates computational resources on regions that actually contain vascular structures, avoiding redundant calculations in blank background areas.

[0128] For each Gaussian kernel in the initial Gaussian kernel set, the spatial position of the voxel point corresponding to the Gaussian kernel is taken as the initial spatial position of the Gaussian kernel.

[0129] The rotation parameters of the Gaussian kernel are initialized to a unit quaternion. This indicates that the Gaussian kernel did not rotate in the initial state.

[0130] The Gaussian kernel that is closest in the initial point cloud Calculate the Gaussian kernel to the nearest neighbor point. The average distance between the neighboring points is used to calculate the initial scale parameters of the Gaussian kernel.

[0131] In this embodiment, the initial spatial position of each Gaussian kernel is directly determined by the three-dimensional coordinates of its corresponding non-empty voxel center point. The initial rotation parameters of all Gaussian kernels are initialized to unit quaternions. This indicates that the Gaussian kernel has not rotated in the initial state. In 3D Gaussian splashing, each Gaussian kernel is no longer a simple point, but an ellipsoid with shape and size in three-dimensional space. The scale parameter is the core parameter used to control the size of this ellipsoid in various dimensions. Specifically, the initial scale parameter of each Gaussian kernel is calculated based on the local point cloud density around the center of the Gaussian kernel. For the center of each Gaussian kernel, i.e., the initial spatial position... Determine its nearest neighbor in the initial point cloud. One neighboring point, Typically, a value of 3 to 5 is used to calculate the value from the initial spatial position. Arrive here Average distance of neighboring points Initial scale parameters The size is set to be the same as this average distance. Proportional, for example, take (0.5 to 1.5 times); of which They represent in Scale parameters in three directions.

[0132] The Kaiming initialization strategy is adopted to initialize the weight parameters of the cluster-shared real-time coding module and the kernel-level MLP with residuals.

[0133] In this embodiment, the initialization of the weight parameters of the cluster-shared real-time coding module and the kernel-level MLP with residuals refers to the fact that all weight parameters of the MLP in the two modules, which are used to predict the Gaussian kernel center decay, adopt the existing Kaiming initialization strategy.

[0134] Clustering algorithms are used to divide all Gaussian kernels in the initial Gaussian kernel set into several spatial clusters, each representing a local blood vessel segment. Cluster labels are then assigned to each Gaussian kernel based on the partitioning results. .

[0135] Based on the cluster label A cluster encoding matrix is ​​initialized for each partitioned spatial cluster. .

[0136] In this embodiment, based on the initial Gaussian kernel spatial distribution, the classic K-Means clustering algorithm is used to divide all Gaussian kernels into multiple spatial clusters, each cluster representing a local blood vessel segment, and a cluster label is assigned to each Gaussian kernel. And initialize the corresponding cluster coding matrix for each blood vessel cluster. The cluster coding matrix is ​​initialized using the existing Xavier method. This step lays the foundation for subsequent clustered temporal modeling.

[0137] The cluster-shared real-time coding (CTEM) module is used to generate a fused temporal feature vector characterizing the dynamic properties of a vascular segment based on the geometric parameters, dynamic decay parameters, and timestamps of the Gaussian kernel, through parallel cluster-level shared coding pathways and kernel-level residual compensation pathways.

[0138] In this embodiment, the CTEM module is divided into two parts: Cluster level (macro): learning a shared, fundamental blood flow dynamic pattern within a blood vessel segment. Kernel level (micro): learning a personalized, fine-tuned adjustment for each Gaussian kernel. For example... Figure 5 As shown, the upper module is the Cluster-wise Shared Encoding Path, the first pillar of the innovation, designed to capture the overall dynamics of the vascular segment. The lower module is the Kernel-wise Residual Compensation Path, the second pillar of the innovation, designed to capture the unique dynamic details of each Gaussian kernel. Specifically, the CTEM module receives the spatial location of the Gaussian kernel. The timestamps corresponding to the training frames sampled from the original DSA sequence. and the cluster label of the Gaussian kernel As input, spatial features are first extracted using 3D hash encoding (H3d). Then, a cluster-sharing mechanism is employed to learn a unified temporal encoding curve for each vessel segment, while kernel-level residual compensation preserves local temporal details. The output of this module is concatenated with the spatial features and then input into the KR-MLP module for further feature fusion.

[0139] The specific content of the cluster-shared real-time coding (CTEM) module is as follows:

[0140] For any Gaussian kernel in a spatial cluster The spatial position of the Gaussian kernel Cluster tags and timestamp As input data, the input data is encoded through parallel cluster-level shared coding pathways and kernel-level residual compensation pathways to generate fusion temporal feature vectors that characterize the dynamic properties of vascular segments.

[0141] The cluster-level shared coding path is as follows: for Gaussian kernels To determine the spatial cluster to which it belongs, obtain the cluster encoding matrix of that spatial cluster. Based on the time basis function, according to the timestamp Generate weight vector The weight vector mentioned above The length of the cluster coding matrix number of rows.

[0142] Based on the weight vector Each element in the cluster encoding matrix The corresponding rows in the vector are scaled, and then all the scaled row vectors are added together to obtain the cluster-level shared coding vector of the spatial cluster.

[0143] The cluster-level shared coding vector is represented as follows:

[0144] (2)

[0145] in This is a cluster-level shared encoding vector; The total number of rows in the cluster coding matrix represents the dimension of the temporal feature decomposition, i.e., the number of frequencies in the temporal coding. Each row stores a learnable feature vector, corresponding to a specific frequency component in the temporal coding. Initialized to 4, indicating there are 4 frequency bands; For the first Row vectors; For the first in the cluster coding matrix OK; Weight vector The first in Each element.

[0146] In this embodiment, the timestamp and cluster tags The input is fed into a time basis function TimeBase, which generates the encoding matrix of the corresponding cluster. Combined with time weights, a unified representation of the cluster at the timestamp is generated. Encoding vector of the global state The combination process in the TimeBase function is implemented through a weighted linear combination, which includes four steps: weight generation, matrix selection, weighted summation of row vectors, and fusion. First, the TimeBase function receives the timestamp. Generate a weight vector The formula for this generation process is as follows:

[0147] (3)

[0148] in This represents the timestamp corresponding to the subtracted DSA sequence; This represents the weight vector corresponding to that timestamp.

[0149] According to cluster label Select the corresponding cluster coding matrix Then, using the weight vector Each element scales the corresponding row of the cluster coding matrix; finally, all scaled row vectors are summed to obtain a single coding vector. This process can be understood as using timestamps... The determined "recipe," or weight, comes from the cluster label. The "raw material library" refers to the feature representations that are selected from the coding matrix to suit the current moment.

[0150] In this embodiment, the kernel-level residual compensation pathway is not processed according to spatial clusters, but is processed independently for each Gaussian kernel.

[0151] The kernel-level residual compensation path is as follows: using a 3D hash encoder to perform Gaussian kernel... spatial location Encode and generate Gaussian kernel The spatial eigenvectors are represented as:

[0152] (4)

[0153] in Represents spatial eigenvectors; Represents the position vector of a single Gaussian kernel; This indicates the use of a 3D hash encoder to measure the Gaussian kernel position vector. The process of encoding; This represents the weight parameter of the 3D hash encoder. The initialization of this parameter is directly given by the hash parameter initialization configuration file. You can manually define the parameter value in the configuration file.

[0154] Gaussian kernel timestamp Perform sine coding to generate a Gaussian kernel. The time feature vector is represented as:

[0155] (5)

[0156] in For time feature vectors, The highest frequency of the sine code indicates that there is One frequency, Essentially, it represents the dimension of temporal feature decomposition, that is, the number of frequencies of temporal encoding. The numerical value and meaning are related to the total number of rows in the cluster coding matrix. The numerical value and meaning are consistent, so here The value is set to 4, which corresponds to 4 frequency bands.

[0157] The spatial and temporal feature vectors are concatenated and input into the residual prediction MLP for learning and prediction, generating a kernel-level residual compensation vector. .

[0158] In this embodiment, as Figure 5 As shown, the kernel-level residual compensation pathway is responsible for learning the unique, subtle temporal dynamic adjustments for each Gaussian kernel. For the input Gaussian kernel spatial location... and the corresponding timestamp Spatial location It is transformed into an efficient spatial feature vector, timestamp, using an existing 3D hash encoder. After passing through an existing sinusoidal encoder, it is transformed into a high-dimensional temporal feature vector to better represent periodic, continuous temporal variations. The spatial and temporal features are then concatenated and input into a lightweight MLP. This MLP learns to predict a residual vector. This is nuclear-grade residual compensation. Finally, cluster-level shared coding vectors and kernel-level residual compensation are applied. They are directly concatenated into a single feature vector output, which serves as the output of the cluster-shared real-time encoding CTEM module.

[0159] Cluster-level shared coding vector and kernel-level residual compensation vector The data are then concatenated to generate a fused temporal feature vector that characterizes the dynamic properties of the vascular segment.

[0160] A kernel-level MLP module with residuals is used to perform feature transformation on the fused temporal feature vector characterizing the dynamic properties of blood vessel segments, generating the center decay value of each Gaussian kernel at a specific time. .

[0161] The kernel-level MLP module with residuals is a lightweight multilayer perceptron containing residual connections, comprising two fully connected units in series. Each fully connected unit includes a fully connected layer (FC) and a ReLU activation function in series. The outputs of the two fully connected units are fused through residual connections and then input into a third fully connected layer (FC3) for dimensionality mapping (mapped to a 1-dimensional scalar) to obtain the timestamp of each Gaussian kernel. center attenuation value .

[0162] like Figure 6As shown, the KR-MLP module introduces a residual connection mechanism on top of the traditional MLP, consisting of three fully connected layers and one residual connection. It preserves low-level feature information through the shortcut path of the residual connection, achieving cross-layer feature fusion and information preservation. This alleviates the gradient vanishing problem in deep network training, improves network training stability and convergence speed, and thus enables more accurate learning of personalized temporal dynamic compensation for each Gaussian kernel. The module outputs the time-stamp of each Gaussian kernel. center attenuation value This process completes the mapping from static geometry to dynamic attenuation. Specifically, the input fused temporal feature vector representing the dynamic characteristics of the vessel segment first passes through the first fully connected layer FC1 and the ReLU activation function to obtain a primary feature representation. This feature simultaneously flows into two processing branches: ① it is directly retained as a residual feature; ② it undergoes deep feature transformation through the second fully connected layer FC2 and the ReLU activation function. Subsequently, the residual feature and the deep transformed feature are fused, and then the final central attenuation value is obtained through a third fully connected layer (FC3) for dimensional mapping. Output.

[0163] In the stage of constructing the 4D radiative Gaussian scene representation, this embodiment introduces the two core innovative modules mentioned above: the Cluster Shared Real-Time Encoding Module (CTEM) and the Kernel-Level MLP with Residue (KR-MLP). Specifically, after initializing the Gaussian kernel and spatial clustering, an optimizable 4D radiative Gaussian model for dynamic DSA reconstruction is further established. This model consists of two parts: geometric parameters and attenuation parameters. The geometric parameters include the position, scale, and rotation properties of the Gaussian kernel; the attenuation parameters include the center attenuation parameter, which is derived from the Cluster Shared Real-Time Encoding Module (CTEM) and the Kernel-Level MLP with Residue (KR-MLP).

[0164] The training process of the vascular 3D reconstruction network based on sparse view dynamic DSA is as follows:

[0165] Build includes The training samples of the sparse view are input into the vascular 3D reconstruction network based on sparse view dynamic DSA for iterative training. In the first iteration, the training samples are input into the vascular 3D reconstruction network based on sparse view dynamic DSA to obtain an initial Gaussian kernel set. From the second iteration onwards, the Gaussian kernel set generated after the previous iteration is used as the optimization object for each iteration. The Gaussian kernel set includes: spatial location... Rotation parameters Scale parameters Cluster coding matrix The network weight parameters of the cluster-shared real-time encoding module and the KR-MLP module are as follows: neural network parameters: weight parameters of the three fully connected layers FC1, FC2 and FC3 in the KR-MLP module; weight parameters of the lightweight MLP in the kernel-level residual compensation path; and weight parameters of the 3D hash encoder.

[0166] For each iteration in the training process, the following procedure is performed:

[0167] Feature extraction and decay prediction: This involves determining the spatial location of the current Gaussian kernel. Cluster tags and timestamp Input cluster-level shared real-time coding module (CTEM). Output cluster-level shared coding path. Nuclear-level residual compensation pathway outputs nuclear-level residual compensation. Features after fusion The input is fed into the KR-MLP module, and the output is each Gaussian kernel. Center decay value at time .

[0168] Furthermore, using a differentiable X-ray rasterization model, the current Gaussian kernel set is rendered into a synthetic DSA image. .

[0169] Determine if the current iteration round has reached the preset maximum number of iterations. If it has, obtain the trained 3D blood vessel reconstruction network based on sparse viewpoint dynamic DSA using the optimized Gaussian kernel set for the current iteration round. If it has not reached the maximum number of iterations, calculate the synthesized DSA image. Compared with the real DSA image under the current sparse view The loss value between the two is used to update the Gaussian kernel set in the current iteration through the backpropagation algorithm.

[0170] This step, performed after a fixed iteration interval, involves adaptively adjusting the distribution and number of Gaussian kernels to optimize their spatial structure. It typically follows parameter optimization and includes the following processes:

[0171] Average decay calculation: In each iteration, update the average decay value of each Gaussian kernel. This average value is the decay value at the center of the Gaussian kernel over multiple iterations. The result is obtained by averaging the sums, and can be expressed as:

[0172] (6)

[0173] Prune: Reducing the average decay value over a long period (multiple iterations). Less than the preset threshold The Gaussian kernel is identified as background noise and removed from the Gaussian kernel set, i.e., pruning.

[0174] Clone: ​​For a small Gaussian kernel in a region of insufficient reconstruction (usually a region of high loss or large gradient), a new Gaussian kernel is generated by copying the Gaussian kernel. The two kernels are in the same location, but the scale is usually reduced to increase local expressive power.

[0175] Split: For large Gaussian kernels that are in over-reconstructed regions (usually regions with large size and large gradient), split the Gaussian kernel into two or more smaller kernels to fit complex geometric boundaries more finely.

[0176] Based on the set of all optimized Gaussian kernels, i.e., the 4D Gaussian kernel, its spatiotemporal parameters include geometric parameters and central decay values. Only the first iteration uses the initial 4D Gaussian kernel; subsequent iterations use the geometric parameters of the optimized Gaussian kernel at specific timestamps. Calculated center attenuation value A well-known differentiable X-ray rasterization model, based on the X-ray attenuation law and widely used in DSA reconstruction techniques based on radiation attenuation fields and Gaussian splashing, is employed to project and render a 4D (spatiotemporal) Gaussian kernel set into a synthetic DSA image. The dynamic attenuation calculation for each Gaussian kernel utilizes clustering information, but the rendering itself is performed on the complete Gaussian kernel set. Subsequently, the loss value between the synthetic DSA image and the ground sparse view image is calculated for gradient optimization. This ground sparse view image is the previously sampled image, a small subset of images uniformly sampled at fixed intervals from the complete DSA scan dataset—specifically, the uniformly sampled image obtained in the first iteration. Frame sparse view.

[0177] It should be noted that, in this embodiment, the input dynamic DSA image sequence used for model training is obtained by sparse view sampling, which is uniformly sampled from the complete sequence. The frames are used as training data; after the model training is completed, the remaining frames that were not used in the training are used as test frames for evaluating the rendering quality of new views or verifying reconstruction performance.

[0178] Will The frame sparse view is input to the trained vascular 3D reconstruction network based on sparse view dynamic DSA, generating the geometric parameters of each Gaussian kernel and the center decay value at all timestamps.

[0179] Calculate the timestamp based on the geometric parameters of each Gaussian kernel and the central decay value across all timestamps. Three-dimensional blood vessel volume .

[0180] The specific details of calculating the three-dimensional blood vessel volume at all time points based on the geometric parameters of each Gaussian kernel and the center attenuation value at all time points are as follows:

[0181] When using the FDK algorithm on the input During the 3D reconstruction of the sparse frame view, the region of interest (ROI) is determined.

[0182] For any non-empty voxel sampling point in the region of interest (ROI) Based on non-empty voxel sampling points Corresponding Gaussian kernel Geometric parameters, calculate non-empty voxel sampling points Spatial Gaussian distribution function This function value is used to represent non-empty voxel sampling points. Weights in the region of interest (ROI).

[0183] For any timestamp According to Gaussian kernel In timestamp center attenuation value Calculate timestamps lower non-empty voxel center point The overall element value.

[0184] After model training is complete, based on the optimized 4D Gaussian kernel, the FDK algorithm is applied to the input in three-dimensional space at each time point. During the 3D reconstruction process using a sparse frame view, the identified Region of Interest (ROI) is voxelized. Specifically, for any non-empty voxel sampling point in the 3D space, the overall voxel value of the center point is calculated based on the center decay values ​​of each Gaussian kernel output by the model at a specific time stamp, combined with its spatial Gaussian distribution function. The spatial Gaussian distribution function, determined by the geometric parameters such as the position, scale, and rotation of the Gaussian kernel, is used to calculate the weight of the non-empty voxel center point in the Gaussian kernel space.

[0185] (7)

[0186] in Indicates the center point of a non-empty voxel. The overall element value.

[0187] timestamp Perform voxel integration on the total voxel values ​​of all non-empty voxel centers to obtain the voxel values ​​at the timestamp. Three-dimensional blood vessel volume .

[0188] In this embodiment, as Figure 7 As shown, by performing the above voxel integration operation, i.e., summing, on all voxels, the three-dimensional blood vessel volume at that moment can be obtained. This is used to represent the spatial voxel distribution of the contrast agent within that time frame, and can reflect the dynamic filling morphology of blood vessels.

[0189] By performing voxelized averaging on the three-dimensional blood vessel volumes at all timestamps, a static 3D blood vessel volume is obtained, which in turn generates a static three-dimensional blood vessel model.

[0190] Based on the geometric parameters of all Gaussian kernels and the center attenuation value at all timestamps, a differentiable X-ray rasterization model is used to render all Gaussian kernels, generating a synthetic DSA image from a new perspective.

[0191] In this embodiment, as Figure 7 As shown, using a trained 3D vascular reconstruction network based on sparse-view dynamic DSA (which includes an optimized 4D radiative Gaussian model and all its parameters), a 2D DSA image observed from a viewpoint not used during training (i.e., a new projection angle) is rendered using differentiable X-ray rasterization at a specified time stamp. This rendered image is referred to as the synthesized new viewpoint.

[0192] In this embodiment, the synthetic DSA image from the new perspective and the static three-dimensional vascular model are two independent results that this solution aims to obtain.

[0193] In this embodiment, quantitative evaluation metrics, such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), are calculated between the synthesized new perspective view and the actual DSA image acquired under that new perspective view to verify the model's generalization ability and reconstruction accuracy.

[0194] 1. Peak Signal-to-Noise Ratio (PSNR): This measures the pixel-level error between a synthetic image and a real image; a higher value indicates better quality. The formula is as follows:

[0195] (8)

[0196] in It is the maximum pixel value of the image; It is the mean square error, defined as:

[0197] (9)

[0198] in and These are the height and width of the image, respectively; and These are synthetic and real DSA images, respectively.

[0199] 2. Structural Similarity Index (SSIM): This index measures the structural similarity of images, considering brightness, contrast, and structural information. A value closer to 1 indicates higher similarity. The formula is as follows:

[0200] (10)

[0201] in and These refer to the synthetic image patch and the real image patch, respectively, which are the corresponding local regions in the synthetic DSA image and the real acquired DSA image; for The average brightness, specifically referring to the composite image patch. The arithmetic mean of all pixel values; for Average brightness, specific value of real image patch The arithmetic mean of all pixel values; for The variance is used to measure the synthetic image patch. The contrast ratio, i.e., the fluctuation range of pixel values; for The variance is used to measure the true image patch. Contrast; for and The covariance is used to measure the structural similarity between two image patches, that is, the degree of linear correlation between them; and These are all very small stability constants, used to avoid the denominator being zero while maintaining the continuity of the function. They are typically set as follows: , , This represents the maximum dynamic range of pixel values, such as 255. and These are all very small constants, such as 0.01.

[0202] Example 2:

[0203] This embodiment presents a three-dimensional vascular reconstruction system based on digital subtraction angiography, such as... Figure 8 As shown, the system includes:

[0204] The image sequence acquisition module is used to acquire two-dimensional digital subtraction angiography (DSA) image sequences of vascular regions obtained through rotational acquisition.

[0205] The registration and subtraction module employs the VoxelMorph non-rigid image registration method based on deep learning to register the two-dimensional digital subtraction angiography (DSA) image sequence frame by frame, and then performs subtraction processing on the registered image sequence to generate the subtracted vascular sequence.

[0206] The uniform sampling module is used to uniformly sample the subtracted vascular sequence to obtain... Frame sparse view.

[0207] The 3D reconstruction module is used to call the trained vascular 3D reconstruction network based on sparse viewpoint dynamic DSA, and to... The frame sparse view is input to the trained vascular 3D reconstruction network based on sparse view dynamic DSA, generating the geometric parameters of each Gaussian kernel and the center decay value at all timestamps.

[0208] The temporal volume generation module is used to calculate the three-dimensional blood vessel volume at all time points based on the geometric parameters of each Gaussian kernel and the center decay value at all time points.

[0209] The static model generation module is used to obtain a static 3D blood vessel volume by voxelizing and averaging the 3D blood vessel volume at all timestamps, and then generate a static 3D blood vessel model.

[0210] The rendering and compositing module is used to render all Gaussian kernels using a differentiable X-ray rasterization model based on the geometric parameters of all Gaussian kernels and the center attenuation value at all timestamps, generating a synthetic DSA image from a new perspective.

[0211] Example 3:

[0212] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the aforementioned method for three-dimensional reconstruction of blood vessels based on digital subtraction angiography.

[0213] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements a three-dimensional vascular reconstruction method based on digital subtraction angiography as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0214] The processor is used to execute all or part of the steps in the three-dimensional vascular reconstruction method based on digital subtraction angiography as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0215] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the three-dimensional vascular reconstruction method based on digital subtraction angiography described in the above embodiments.

[0216] Example 4:

[0217] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0218] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the three-dimensional vascular reconstruction method based on digital subtraction angiography described in the various embodiments of this application.

[0219] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned method for three-dimensional vascular reconstruction based on digital subtraction angiography.

[0220] Example 5:

[0221] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned method for three-dimensional vascular reconstruction based on digital subtraction angiography.

[0222] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0223] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0224] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A method for three-dimensional reconstruction of blood vessels based on digital subtraction angiography, characterized by, The method comprises the following steps: obtaining a two-dimensional digital subtraction angiography (DSA) image sequence of a blood vessel region collected through rotation; using a VoxelMorph non-rigid image registration method based on deep learning to perform frame-by-frame registration on the two-dimensional digital subtraction angiography (DSA) image sequence, and performing subtraction processing on the registered image sequence to generate a subtracted blood vessel sequence; The blood vessel sequence after subtraction is uniformly sampled to obtain Frame sparse view; constructing a blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA and training to obtain a trained blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA; Will The frame sparse view input is trained into the blood vessel three-dimensional reconstruction network based on the sparse view dynamic DSA, to generate the geometric parameters of each Gaussian kernel and the central attenuation value at all time stamps; calculating the three-dimensional blood vessel volume at all time stamps according to the geometric parameters of each Gaussian kernel and the central attenuation value at all time stamps; performing voxel-wise averaging on the three-dimensional blood vessel volume at all time stamps to obtain a static 3D blood vessel volume, and further generating a static three-dimensional blood vessel model; based on the geometric parameters of all Gaussian kernels and the central attenuation value at all time stamps, rendering all Gaussian kernels through a differentiable X-ray rasterization model to generate a synthetic DSA image at a new view angle.

2. The method of claim 1, wherein the method is based on digital subtraction angiography. The specific content of the step of obtaining a two-dimensional digital subtraction angiography (DSA) image sequence of a blood vessel region collected through rotation is as follows: without injecting contrast medium, a set of continuous two-dimensional background image sequences is obtained through first rotation collection along a preset rotation trajectory, denoted as a mask sequence; after injecting contrast medium into the blood vessel, a second rotation collection is performed along the same rotation trajectory as in the first rotation collection to obtain a set of continuous two-dimensional dynamic contrast sequences, denoted as a filling sequence; The acquisition parameters, the view angle, and the frame rate of the second rotation acquisition are consistent with those of the first rotation acquisition, so as to ensure that the first frame image in the filling sequence and the first frame image in the mask sequence are consistent. The first frame image in the filling sequence and the first frame image in the mask sequence are consistent. The first frame image in the filling sequence and the first frame image in the mask sequence are consistent. the mask sequence and the filling sequence are taken as the two-dimensional digital subtraction angiography (DSA) image sequence of the blood vessel region.

3. The method of claim 2, wherein the method further comprises: The specific content of the step of using a VoxelMorph non-rigid image registration method based on deep learning to perform frame-by-frame registration on the two-dimensional digital subtraction angiography (DSA) image sequence, and performing subtraction processing on the registered image sequence to generate a subtracted blood vessel sequence is as follows: a plurality of groups of paired mask frames and filling frames are obtained, each group of paired mask frames and filling frames is spliced in the channel dimension, and a training set is constructed using the double-channel tensor obtained by splicing; for any double-channel tensor in the training set, the double-channel tensor is input into a convolutional neural network to output a two-dimensional displacement vector field (DVF); based on the two-dimensional displacement vector field (DVF), a spatial transformation network is used to register the filling frame corresponding to the double-channel tensor to obtain a registered filling image; a joint loss of the convolutional neural network is constructed, and the network parameters of the convolutional neural network are optimized by minimizing the joint loss; For any timestamp extracting a corresponding mask frame and a filling frame from the two-dimensional digital subtraction angiography (DSA) image sequence and a filling frame are spliced in the channel dimension to obtain a dual-channel tensor; the double-channel tensor is input into the optimized convolutional neural network to output a two-dimensional displacement vector field (DVF); Based on a two-dimensional displacement vector field (DVF), a spatial transformation network is used to register a fill frame to obtain a registered fill image ; The log domain pixel values of the registered fill image and the mask frame are calculated respectively, and the registered fill image is subtracted from the mask frame pixel by pixel to obtain the time-stamped subtraction blood vessel image; a subtracted blood vessel sequence is generated according to the subtracted blood vessel images at all time stamps.

4. The method of claim 3, wherein the method is based on digital subtraction angiography. The blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA comprises: The preliminary 3D reconstruction module is used to perform FDK algorithm on the input. The sparse view of the frame is used for 3D reconstruction to generate a 3D blood vessel volume model composed of several individual pixels. a 4D Gaussian initialization module for non-empty voxel sampling on the three-dimensional blood vessel volume model, and constructing an initial Gaussian kernel set according to the voxel points obtained by sampling, and initializing the geometric parameters and dynamic attenuation parameters of each Gaussian kernel in the set; The clustered shared real-time encoding module is configured to generate a fusion time sequence feature vector representing dynamic characteristics of the vessel segment according to geometric parameters, dynamic attenuation parameters and time stamps of the Gaussian kernel, through a parallel cluster-level shared encoding channel and a kernel-level residual compensation channel. The kernel-level MLP module with residual is configured to perform feature transformation on the fusion time sequence feature vector representing the dynamic characteristics of the vessel segment, to generate a central attenuation value of each Gaussian kernel at a specific time.

5. The method of claim 4, wherein the method further comprises: Geometric parameters of the Gaussian kernel, including: spatial position , rotation parameter and scale parameter ; dynamic attenuation parameter, including: cluster encoding matrix ; The specific content of the 4D Gaussian initialization module is as follows: traversing all the voxel points in the three-dimensional blood vessel volume model, comparing each voxel point with a preset voxel threshold value and retaining all the voxel points with a voxel value greater than ​ Each reserved voxel point is taken as a Gaussian kernel to construct an initial Gaussian kernel set. For each Gaussian kernel in the initial Gaussian kernel set, the spatial position of the voxel point corresponding to the Gaussian kernel is taken as the initial spatial position of the Gaussian kernel. initializing the rotation parameters of the Gaussian kernel to unit quaternions , indicating that in the initial state, the Gaussian kernel has not been rotated; The Gaussian kernel that is closest in the initial point cloud Calculate the Gaussian kernel to the nearest neighbor point. The average distance between the neighboring points is used to calculate the initial scale parameters of the Gaussian kernel. The Kaiming initialization strategy is adopted to initialize the weight parameters of the clustered shared real-time encoding CTEM module and the kernel-level MLP module with residual. All the Gaussian kernels in the initial Gaussian kernel set are divided into several spatial clusters by using a clustering algorithm, and each spatial cluster represents a local vessel segment, and then each Gaussian kernel is assigned a cluster label according to the division result ; based on the cluster label initializing a cluster encoding matrix for each divided space cluster .

6. The method of claim 5, wherein the method further comprises: The specific content of the clustered shared real-time encoding module is as follows: For all Gaussian kernels in any spatial cluster The spatial position of the Gaussian kernel , the cluster label and the timestamp are taken as input data, and the input data is encoded through a parallel cluster-level shared encoding channel and a kernel-level residual compensation channel to generate a fused time sequence feature vector representing the dynamic characteristics of the blood vessel segment. The cluster-level shared encoding path is: for a Gaussian kernel The spatial cluster to which the spatial cluster belongs, and obtain a cluster encoding matrix of the spatial cluster ; based on a time basis function, from a timestamp generating a weight vector ; wherein the weight vector has a length of a number of rows of a cluster encoding matrix ​ According to the weight vector each element in the cluster encoding matrix The corresponding row in the cluster encoding matrix is scaled, and all the scaled row vectors are added to obtain the cluster-level shared encoding vector of the space cluster. The kernel-level residual compensation path is as follows: using a 3D hash encoder to perform Gaussian kernel... spatial location Encode and generate Gaussian kernel Spatial eigenvectors; Gaussian kernel timestamp Perform sine coding to generate a Gaussian kernel. Time feature vector; Gaussian kernel The spatial and temporal feature vectors are concatenated and input into the residual prediction MLP for learning and prediction, generating a Gaussian kernel. Kernel-level residual compensation vector ; Cluster-level shared encoding vectors and core-level residual compensation vectors The splicing is performed to generate a fusion time sequence feature vector representing dynamic characteristics of the blood vessel segment.

7. The method of claim 6, wherein the method further comprises: The residual core-level MLP module is a lightweight multi-layer perceptron including a residual connection, including two full connection units in series; the full connection unit includes a full connection layer FC and a ReLU activation function in series; the outputs of the two full connection units are fused through the residual connection and then input into a third full connection layer FC3 for dimension mapping to obtain a center decay value of each Gaussian kernel at a timestamp . .

8. The method of claim 7, wherein the method is based on digital subtraction angiography. The training process of the blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA is as follows: constructing a training sample comprising frame sparse view, inputting the training sample into a blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA for iterative training; in the first round of iteration, inputting the training sample into the blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA to obtain an initial Gaussian kernel set; From the second iteration, the Gaussian kernel set generated after the previous iteration is taken as the optimization object of each iteration. The Gaussian kernel set comprises: a spatial position , a rotation parameter , a scale parameter , a cluster encoding matrix , and network weight parameters of the cluster sharing real-time encoding module and the KR-MLP module; For each iteration in the training process, the following process is performed: Based on the current set of Gaussian kernels, the center attenuation values of all Gaussian kernels are predicted using the clustered shared real-time encoding module and the KR-MLP module, and then the current set of Gaussian kernels is rendered into a synthetic DSA image through the differentiable X-ray rasterization model ; determine whether the current iteration round reaches the preset maximum iteration number, if yes, obtain the trained blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA according to the optimized Gaussian kernel set in the current iteration round, if not, calculate the synthetic DSA image and the real DSA image in the current sparse view between the loss value, and update the Gaussian kernel set in the current iteration round by using the loss value through the back propagation algorithm. After completing the fixed iteration interval, the pruning, cloning or splitting operation is performed on the Gaussian kernel set under the current iteration round, to complete the adaptive adjustment of the Gaussian kernel set under the current iteration round.

9. The method of claim 8, wherein the method further comprises: The specific content of calculating the three-dimensional blood vessel volume at all time stamps according to the geometric parameters of each Gaussian kernel and the central attenuation value at all time stamps is as follows: When using the FDK algorithm on the input During the 3D reconstruction of sparse frame views, the region of interest (ROI) is determined. For any non-empty voxel sampling point in the region of interest (ROI) Based on non-empty voxel sampling points Corresponding Gaussian kernel Geometric parameters, calculate non-empty voxel sampling points Spatial Gaussian distribution function This function value is used to represent non-empty voxel sampling points. Weights in the region of interest (ROI); For any timestamp , according to a Gaussian kernel At the center of the timestamp , the attenuation value , the total voxel value of the non-empty voxel center point under the timestamp is calculated ; timestamp The voxel integration operation is performed on the total voxel values of all non-empty voxel center points below the timestamp to obtain the three-dimensional blood vessel volume at the timestamp 10. A digital subtraction angiography-based blood vessel three-dimensional reconstruction system for implementing the digital subtraction angiography-based blood vessel three-dimensional reconstruction method according to any one of claims 1 to 9, characterized by, The system comprises: An image sequence acquisition module is configured to acquire a two-dimensional digital subtraction angiography (DSA) image sequence of a blood vessel region obtained by rotation acquisition; A registration subtraction module is configured to perform frame-by-frame registration on the two-dimensional digital subtraction angiography (DSA) image sequence by using a deep learning-based VoxelMorph non-rigid image registration method, and to perform subtraction processing on the registered image sequence to generate a subtracted blood vessel sequence; The uniform sampling module is configured to uniformly sample the blood vessel sequence after subtraction to obtain Frame sparse view; a three-dimensional reconstruction module, configured to call a trained blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA, and input the sparse view dynamic DSA to the trained blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA to obtain a three-dimensional reconstruction result of the blood vessel in the sparse view dynamic DSA. frame sparse view input is input to the trained blood vessel three-dimensional reconstruction network based on sparse view dynamic DSA to generate geometric parameters of each Gaussian kernel and central attenuation values at all time stamps. A time sequence volume generation module is configured to calculate the three-dimensional blood vessel volume at all time stamps according to the geometric parameters of each Gaussian kernel and the central attenuation value at all time stamps. A static model generation module is configured to obtain a static 3D blood vessel volume by voxelizing and averaging the three-dimensional blood vessel volumes at all time stamps, and to generate a static three-dimensional blood vessel model. A rendering synthesis module is configured to perform rendering on all Gaussian kernels by using a differentiable X-ray rasterization model based on the geometric parameters of all Gaussian kernels and the central attenuation values at all time stamps, to generate a synthesized DSA image under a new view angle.

Images