Calcification topology prior-based and graph-voxel coupled coronary cta image generation method and system

Abstract

The application discloses a kind of based on calcification topological priori and graph-voxel coupling coronary artery CTA image generation method and system, specifically is: collection electrocardiogram control heart CT image and coronary artery CTA image, and three-dimensional elastic registration is carried out to CT image;Extract electrocardiogram phase information from the header file of original DICOM data, and encode as high-dimensional electrocardiogram phase vector;Extraction calcification point set is carried out on CT image by threshold segmentation;Connect the calcification point set based on topological path generation algorithm, form three-dimensional continuous path, and generate topological probability graph reflecting the probability distribution of blood vessel center line;The registered CT image and topological probability graph are input into the graph-voxel coupling generation network trained in advance, and inject electrocardiogram phase vector to carry out conditional control, generate target coronary artery CTA image.The application utilizes contrast agent-free electrocardiogram control heart CT image, that is, can accurately generate virtual coronary artery CTA image.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, and in particular to a method and system for generating coronary CTA images based on calcification topological prior and graph-voxel coupling. Background Technology

[0002] Coronary artery disease (CAD) is one of the leading causes of death and disability worldwide. Early and accurate diagnosis is crucial for improving patient outcomes. Coronary CTA is currently the non-invasive "gold standard" for diagnosing coronary artery disease. It requires the injection of an iodine-containing contrast agent to clearly visualize narrowing, plaque, and other conditions in the coronary arteries. However, the contrast agent can cause allergic reactions, kidney damage, and the procedure is relatively expensive.

[0003] Cardiac CT plain scan is commonly used for coronary artery calcium score detection. It does not require contrast agent injection, has a fast scanning speed, low radiation dose, and low risk. However, it can only display calcified plaques and cannot assess luminal stenosis and non-calcified plaques. Currently, the routine procedure for obtaining coronary CTA images is as follows: (1) First, perform a cardiac CT plain scan: for calcium score calculation and localization; (2) Then, perform a coronary CTA scan: after contrast agent injection, the scan is performed under ECG gating to obtain clear vascular images. This procedure is the gold standard in clinical practice, but it requires the above two steps, which has the following disadvantages: (1) Radiation dose: Coronary CTA scan itself requires a high radiation dose to ensure image quality. Performing another CTA scan after a plain scan significantly increases the cumulative radiation dose to the patient; (2) Iodine contrast agent: Some patients have allergic reactions to iodine contrast agents, which may be life-threatening in severe cases; Contrast agents may cause contrast agent nephropathy, which poses a risk to patients with existing renal insufficiency; Iodine contrast agents are contraindicated for patients with uncontrolled hyperthyroidism; (3) Process efficiency: Two scans prolong the total examination time and reduce the efficiency of medical equipment turnover; (4) High cost: Contrast agents themselves have costs, and the entire examination process is relatively complex, requiring a large investment of manpower and resources.

[0004] In recent years, the rapid development of deep learning technology has led to some research attempts in academic fields to utilize artificial intelligence for medical image generation or enhancement. For example, low-dose CT reconstruction uses deep learning models to denoise low-dose CT images, generating images with quality similar to conventional-dose CT images. Image-to-image translation, such as using CycleGAN to convert MRI images into CT images, or "pseudo-enhanced" non-enhanced CT images into contrast-enhanced images, provides new ideas for the field. However, existing AI research has a shortcoming: it fails to fully explore and utilize ECG gating information and the unique anatomical structure and dynamic characteristics of coronary arteries. This results in "pseudo-CTA" images with severe distortion in details such as vessel edges, stenosis degree, and plaque composition, failing to meet the accuracy required for clinical diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating coronary CTA images based on calcification topological prior and graph-voxel coupling. This method eliminates the need for additional contrast agent injection and a second CTA scan. It utilizes conventionally acquired, ECG-gated cardiac CT plain scan images and a deep learning model to accurately generate coronary CTA images suitable for clinical diagnosis.

[0006] The technical solution to achieve the purpose of this invention is: a method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling, comprising the following steps:

[0007] S1. Construct a multimodal spatiotemporally aligned cardiac CT dataset: Collect ECG-gated plain CT images and coronary CTA images of the same patient, and perform three-dimensional elastic registration of the plain CT images based on the CTA images; extract ECG phase information from the header file of the original DICOM data and encode it into a high-dimensional ECG phase vector;

[0008] S2. Constructing a calcification topology prior map: On the registered plain CT image, calcification point set is extracted by threshold segmentation; the calcification point set is connected based on the topology path generation algorithm to form a three-dimensional continuous path, and a topology probability map reflecting the probability distribution of the vessel centerline is calculated and generated.

[0009] S3. Generate CTA images using a graph-voxel coupling network: Input the registered plain CT image and the topological probability map into a pre-trained graph-voxel coupling network, and inject the ECG phase vector for conditional control to generate the target coronary artery CTA image.

[0010] The graph-voxel coupling generation network adopts a dual-branch heterogeneous architecture, including:

[0011] The voxel-Transformer branch is used to extract global context voxel features from plain CT images;

[0012] The topology-graph reasoning branch is used for feature reasoning and propagation on a vascular topology consisting of calcifications and centerline points;

[0013] A bidirectional feature projection module is used to realize bidirectional feature mapping and fusion between the voxel-Transformer branch and the topology-graph reasoning branch;

[0014] The conditional decoder is used to decode and output coronary CTA images based on fusion features and ECG phase vectors.

[0015] A coronary CTA image generation system based on calcification topology prior and graph-voxel coupling is disclosed. This system implements the aforementioned coronary CTA image generation method based on calcification topology prior and graph-voxel coupling. The system includes a multimodal data preprocessing module, a calcification topology prior construction module, and a graph-voxel coupling generation inference module, wherein:

[0016] The multimodal data preprocessing module is configured to parse the acquired DICOM data, complete the three-dimensional elastic registration of the plain CT images, and extract the ECG phase vector.

[0017] The calcification topology prior construction module is configured to automatically identify calcification regions in plain CT images and construct a three-dimensional vascular centerline path and corresponding adjacency matrix based on graph theory algorithms.

[0018] The graph-voxel coupling generation inference module loads the trained graph-voxel coupling generation network model, runs the voxel-transformer branch to process voxel data and the topology-graph inference branch to process topology data in parallel, and realizes feature interaction through the bidirectional feature projection module. Finally, it directly outputs virtual coronary CTA images through the ECG-modulated conditional decoder.

[0019] An electronic device includes a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to realize the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling.

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for generating coronary CTA images based on calcification topological priors and graph-voxel coupling.

[0021] Compared with the prior art, the significant advantages of this invention are:

[0022] (1) Precise anatomical structure, preventing breakage and drift: This invention utilizes the graph-voxel coupling generation network GVC-Net and calcification topology prior to force the generated blood vessels to strictly connect to real calcification points, effectively solving the problem of easy breakage and misplacement of blood vessels generated by traditional AI, and ensuring the high realism of the anatomical direction of blood vessels.

[0023] (2) Extremely safe and without contrast agents: The use of iodine contrast agents is completely avoided, and the risk of allergic reactions and contrast agent nephropathy (CIN) is completely eliminated, so that patients with renal insufficiency, hyperthyroidism and allergic constitution can also safely undergo vascular assessment.

[0024] (3) Radiation dose is significantly reduced: patients only need to undergo one low-dose plain scan and do not need to undergo a second high-dose CTA scan, resulting in an overall radiation dose reduction of more than 50%;

[0025] (4) Clear images without artifacts: Combining ECG phase modulation and frequency-spatial domain joint constraints effectively suppresses cardiac motion artifacts, resulting in sharp edges and clear textures of the generated blood vessel walls, meeting the needs of clinical screening and diagnosis.

[0026] (5) Cost reduction and efficiency improvement: It simplifies the clinical examination process, eliminates the need for injection consumables, shortens the examination time, significantly reduces medical costs and improves equipment turnover efficiency. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart of the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling of the present invention.

[0028] Figure 2 This is a schematic diagram illustrating the principle of the coronary CTA image generation system based on calcification topological prior and graph-voxel coupling of the present invention.

[0029] Figure 3 This is a diagram of the global context branch network structure of the voxel-Transformer of this invention. Detailed Implementation

[0030] This invention provides a method and system for generating coronary CTA images based on calcification topological prior and graph-voxel coupling, aiming to directly generate virtual coronary CT angiography (CTA) images from contrast-free ECG-gated non-contrast cardiac CT images. To achieve this goal, this invention proposes a "Graph-Voxel Coupled Generative Network (GVC-Net)". This network is not limited to a single convolution operation, but constructs a heterogeneous feature interaction space. It uses a Transformer to capture the global geometric features of the heart in the voxel domain, and simultaneously uses a Graph Neural Network (GCN) to perform explicit reasoning on the vascular tree in the topological domain. A bidirectional projection mechanism is used to achieve deep fusion of the two, thereby accurately generating continuous and clear coronary vascular trees.

[0031] This invention provides a method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling. This method utilizes deep learning to generate coronary CTA images from ECG-gated cardiac CT scans. The specific implementation process is as follows: Figure 1 As shown, it includes the following steps:

[0032] S1. Construct a multimodal spatiotemporally aligned cardiac CT dataset: Collect ECG-gated plain CT images and coronary CTA images of the same patient, and perform three-dimensional elastic registration of the plain CT images based on the CTA images; extract ECG phase information from the header file of the original DICOM data and encode it into a high-dimensional ECG phase vector;

[0033] S2. Constructing a calcification topology prior map: On the registered plain CT image, calcification point set is extracted by threshold segmentation; the calcification point set is connected based on the topology path generation algorithm to form a three-dimensional continuous path, and a topology probability map reflecting the probability distribution of the vessel centerline is calculated and generated.

[0034] S3. Generate CTA images using a graph-voxel coupling network: Input the registered plain CT image and the topological probability map into a pre-trained graph-voxel coupling network, and inject the ECG phase vector for conditional control to generate the target coronary artery CTA image.

[0035] The graph-voxel coupling generation network adopts a dual-branch heterogeneous architecture, including:

[0036] The voxel-Transformer branch is used to extract global context voxel features from plain CT images;

[0037] The topology-graph reasoning branch is used for feature reasoning and propagation on a vascular topology consisting of calcifications and centerline points;

[0038] A bidirectional feature projection module is used to realize bidirectional feature mapping and fusion between the voxel-Transformer branch and the topology-graph reasoning branch;

[0039] The conditional decoder is used to decode and output coronary CTA images based on fusion features and ECG phase vectors.

[0040] As a specific example, in step S1, the ECG phase information is encoded into a high-dimensional ECG phase vector, as follows:

[0041] Extract the percentage of the ECG RR interval at the scan time from the header file of the original DICOM data. ;

[0042] The percentage of the RR interval Through an embedding layer containing a multilayer perceptron, it is mapped to a high-dimensional ECG phase vector of a set length. The mapping relationship is expressed as:

[0043]

[0044] in, This represents a multilayer perceptron; This indicates a phase embedding operation, using sinusoidal position coding.

[0045] As a specific example, step S2 specifically includes:

[0046] S21. Calcification point extraction: Apply adaptive threshold segmentation to plain CT images and extract calcification points with gray values ​​higher than a set threshold. The pixels constitute a three-dimensional calcification point set. :

[0047]

[0048] in, Represents the spatial domain of a three-dimensional image. Represents the position coordinates in the image. Indicates the position coordinates of a plain CT image. The grayscale value at that location;

[0049] S22. Topological Probability Graph Generation: Using the minimum spanning tree algorithm or the geodesic distance transformation algorithm, the set of calcified points is connected to generate a continuous three-dimensional topological path. ;

[0050] Calculate the topological probability graph The formula is as follows:

[0051]

[0052] in, In voxel coordinates, Voxel representation to path Euclidean distance, This is the attenuation coefficient.

[0053] As a specific example, in step S3, the voxel-Transformer branch of the graph-voxel coupling generation network is configured to process voxel data from plain CT images to capture the geometric structure and long-range dependencies of the heart, specifically:

[0054] Using a 3D Swin Transformer as the backbone network, after plain CT images are input into this network, self-attention is calculated within non-overlapping local windows through a moving window attention mechanism to establish global contextual associations between voxels and generate multi-scale voxel feature maps containing anatomical semantics. .

[0055] As a specific example, the topology-graph inference branch of the graph-voxel coupled generation network in step S3 is configured to use a graph convolutional network to process the sparse topology graph and explicitly infer the connectivity of the vascular tree in non-Euclidean space, as follows:

[0056] Extract calcification points and centerline points from the topological probability graph obtained in step 2. Use the extracted calcification points and centerline points as graph nodes V, and the topological connections between nodes as edges E to construct a vascular topological graph. ;

[0057] Graph Convolutional Network (GCN) is used to propagate and aggregate node feature information along the blood vessel topology. The inter-layer propagation formula is as follows:

[0058]

[0059] in, , Table 1 Layer input and output node feature matrices; Indicates the first The weight matrix of the layer training is used to perform linear transformations on the features; Represents a nonlinear activation function; This represents the adjacency matrix with added self-loops. The original adjacency matrix describing the connectivity relationships between blood vessel nodes. It is the identity matrix; Then it is The degree matrix is ​​used to normalize features during the convolution process.

[0060] As a specific example, in step S3, the bidirectional feature projection module of the graph-voxel coupling generation network performs bidirectional feature mapping to obtain a hybrid feature map, specifically including:

[0061] Voxel to graph projection: Based on the 3D coordinates of the graph nodes, the corresponding feature values ​​are sampled from the voxel feature map output by the voxel-Transformer branch through trilinear interpolation and added to the graph node features;

[0062] Graph-to-voxel projection: Projecting the graph node features output by the topology-graph inference branch. The feature map is rendered into a three-dimensional voxel space using Gaussian sputtering technology, and then fused with the feature map of the voxel-Transformer branch.

[0063] As a specific example, the conditional decoder of the graph-voxel coupled generation network in step S3 contains multiple residual convolutional modules, and an adaptive instance normalization layer is embedded in each residual convolutional module.

[0064] The adaptive instance normalization layer dynamically adjusts the mean and variance of the mixed feature map based on the input ECG phase vector, so that the generated coronary artery morphology corresponds to the cardiac cycle phase at the scan time.

[0065] As a specific example, prior to step S3, a step of training the graph-voxel coupling generation network is also included, specifically:

[0066] Construct a training set that includes plain CT images, topological probability maps, ECG phase vectors, and real CTA images;

[0067] The graph-voxel coupled generation network is trained using a composite loss function. for:

[0068]

[0069] in, Indicating resistance to loss, Indicates L1 loss, This represents the frequency domain consistency loss. Represents graph topology loss. , , They represent , , The weighting coefficients.

[0070] As a specific example, the training process of the graph-voxel coupled generative network adopts a phased strategy:

[0071] In the first stage, the parameters of the voxel-Transformer branch are frozen, and only the topology-graph inference branch is trained to learn the correct vascular topology.

[0072] In the second stage, all network parameters are unfrozen, and the composite loss function is used to jointly optimize and train the overall network until the similarity index of blood vessel topology on the validation set converges, thus obtaining the final graph-voxel coupled generative network.

[0073] This invention also provides a coronary CTA image generation system based on calcification topology prior and graph-voxel coupling. This system is used to implement the aforementioned coronary CTA image generation method based on calcification topology prior and graph-voxel coupling. The system includes a multimodal data preprocessing module, a calcification topology prior construction module, and a graph-voxel coupling generation inference module, wherein:

[0074] The multimodal data preprocessing module is configured to parse the acquired DICOM data, complete the three-dimensional elastic registration of the plain CT images, and extract the ECG phase vector.

[0075] The calcification topology prior construction module is configured to automatically identify calcification regions in plain CT images and construct a three-dimensional vascular centerline path and corresponding adjacency matrix based on graph theory algorithms.

[0076] The graph-voxel coupling generation inference module loads the trained graph-voxel coupling generation network model, runs the voxel-transformer branch to process voxel data and the topology-graph inference branch to process topology data in parallel, and realizes feature interaction through the bidirectional feature projection module. Finally, it directly outputs virtual coronary CTA images through the ECG-modulated conditional decoder.

[0077] The present invention also provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to realize the method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling.

[0078] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for generating coronary CTA images based on calcification topological priors and graph-voxel coupling.

[0079] The coronary CTA image generation method based on calcification topological prior and graph-voxel coupling described in this invention includes the following key technologies:

[0080] (1) Guided generation mechanism for constructing vascular topology prior maps based on calcification points.

[0081] The present invention addresses the issue of invisible soft tissue and blood vessels in plain scan images by not directly allowing the network to make blind predictions. Instead, it first extracts high-density calcification points using an adaptive threshold, treating them as "anchor points." Then, using graph theory algorithms (such as Minimum Spanning Tree (MST) or Fast Proceedings), it constructs "potential vascular centerline paths" connecting these calcification points in three-dimensional space and generates a topological probability graph. As a strong anatomical constraint input for the network.

[0082] Existing technology implementation: Existing technologies typically adopt an end-to-end "black box" learning mode, directly inputting the raw plain scan image into the generative network (such as CycleGAN or Pix2Pix), relying entirely on the network to implicitly learn the blood vessel distribution pattern.

[0083] The difference between the two is that existing technologies are prone to causing blood vessel position drift or generating false blood vessel structures (illusions) in soft tissue areas with no calcification or low contrast; while this invention transforms the medical anatomical a priori principle (blood vessels must be connected to calcification points) into a clear mathematical constraint, forcing the network to generate blood vessels in anatomically reasonable locations, fundamentally ensuring the accuracy of blood vessel orientation.

[0084] (2) Graph-Voxel Coupled Generative Network (GVC-Net) architecture.

[0085] The present invention implements a dual-branch heterogeneous network architecture. One branch utilizes a 3D SwinTransformer to process dense voxels, capturing the global geometric context of the heart; the other branch utilizes a Graph Convolutional Network (GCN) to process a sparse topological graph, explicitly inferring the connectivity of the vascular tree in non-Euclidean space. More importantly, a bidirectional feature projection module (V2G / G2V) is designed to achieve deep interaction and fusion of voxel texture information and topological structure information.

[0086] Existing technology implementation: Most existing technologies use a single convolutional neural network architecture (such as 3D U-Net or ResNet) to try to process large areas of myocardial background and fine linear vascular structures simultaneously with the same set of convolutional kernels.

[0087] The difference between the two is that the pure convolutional architecture of the existing technology is difficult to take into account both global semantics and local elongated structures, which can easily lead to vascular rupture; the heterogeneous architecture of the present invention achieves "divide and conquer" and "complementary advantages". Transformer is good at looking at the whole, GCN is good at connecting breakpoints, and bidirectional projection achieves the unification of the two, thereby accurately generating continuous and complete coronary artery vascular trees.

[0088] (3) Optimization strategy of dual-domain constraint in frequency domain and spatial domain.

[0089] The implementation scheme of this invention: In the loss function for model training, this invention not only constrains the consistency of the pixel space, but also introduces frequency domain consistency loss. That is, a three-dimensional fast Fourier transform (FFT) is performed on the generated image and the real image to calculate the difference between the two in amplitude spectrum and phase spectrum, forcing the model to learn the high-frequency edge features of the blood vessel wall.

[0090] Existing technology implementation: Existing technologies typically use only pixel-level mean square error (MSE) or L1 loss, or rely solely on the adversarial loss of the discriminator.

[0091] The difference between the two is that using only spatial domain loss will result in an "averaging" effect in the generated image, which manifests as blurred blood vessel edges and loss of texture, making it unusable for clinical stenosis measurement; the frequency domain constraint strategy of this invention forces the model to recover high-frequency components, resulting in sharp blood vessel wall edges and high contrast in the generated virtual CTA image, significantly improving the image clarity and diagnostic value.

[0092] (4) Adaptive feature modulation driven by ECG phase vector.

[0093] The present invention implements the following: The present invention extracts the percentage of ECG RR intervals at the scan time from the original DICOM data, maps it into a high-dimensional phase vector through position encoding, and injects it into each layer of the decoder using an adaptive instance normalization (AdaIN) layer to dynamically adjust the mean and variance of the feature map.

[0094] Existing technical implementation: Existing image generation models typically ignore electrocardiogram (ECG) information, or simply concatenate the temporal phase as a simple scalar to the input layer, or even completely disregard ECG information, assuming the heart is static.

[0095] The key difference lies in the fact that the heart is a rapidly beating organ, and the vascular morphology varies significantly across different phases (diastole / systole). Images generated by existing technologies often contain motion artifacts or morphological mismatches; the depth modulation mechanism of this invention endows the model with "physiological perception" capabilities, enabling it to accurately generate vascular morphology corresponding to the cardiac cycle based on electrocardiogram signals, effectively suppressing motion artifacts.

[0096] (5) Integrated non-invasive coronary artery virtual imaging system.

[0097] This invention protects a complete, modular intelligent medical imaging system. Logically, this system integrates a multimodal data preprocessing module (responsible for ECG analysis and elastic registration), a calcification topology prior construction module (responsible for converting image features into graph theory constraints), and a graph-voxel coupling generation and inference module (responsible for running GVC-Net for end-to-end generation). As a complete solution, this system can directly connect to the hospital PACS network, read plain scan DICOM data, and output virtual CTA images without manual intervention.

[0098] Existing technology implementation: Existing medical imaging workstations typically only have "passive display" or "basic post-processing" functions (such as MPR, VR reconstruction). To obtain CTA images, a second physical contrast-enhanced scan must be performed using CT equipment hardware (involving high-pressure injectors and contrast agents). The workstation itself does not have the ability to generate angiography data "out of thin air" from plain scan data.

[0099] The difference lies in the operating mode: the existing system operates on a "physical acquisition-passive display" model, heavily reliant on contrast agents and hardware scanning sequences; the operating mode of the system of this invention is "data-driven-active generation," endowing ordinary CT workstations with the new function of "virtual imaging." It changes the attributes of medical equipment, enabling it to provide advanced diagnostic information solely from plain scan data, fundamentally reconstructing the clinical examination process for coronary heart disease (from "dual scans" to "single scan + intelligent calculation").

[0100] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description is provided in conjunction with embodiments and... Figures 1-3 The present invention will be further described in detail using coronary artery tree reconstruction as an example. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0101] Example 1

[0102] This embodiment provides a method for generating coronary CTA images based on calcification topological priors and graph-voxel coupling, the process of which is as follows: Figure 1 As shown, the steps are as follows:

[0103] Step 1: Construct a multimodal, spatiotemporally aligned cardiac CT dataset;

[0104] Step 1.1: Collect ECG-gated non-contrast CT (NCCT) and coronary CTA images of the same patient. Using CTA as a reference, perform three-dimensional elastic registration on the NCCT to eliminate differences in respiratory motion and ensure spatial consistency of anatomical structures.

[0105] In this specific implementation, 1000 retrospective clinical data were selected. First, the DICOM images were preprocessed, and all images were resampled to an isotropic resolution of 0.5mm×0.5mm×0.5mm. The CT values ​​(HU) were truncated to the range of [-200, 800] and then normalized to the interval of [0, 1].

[0106] During the registration process, a B-spline free-form deformation algorithm based on mutual information as the similarity metric is used. A three-level multi-resolution pyramid strategy is set, with grid control point spacing of [32, 16, 8] mm and a maximum number of iterations of 100, thereby achieving sub-pixel level 3D elastic alignment.

[0107] Step 1.2: ECG Phase Vector Extraction. Extract the percentage of the ECG RR interval at each scan time from the header file of the original DICOM data. It is mapped to a high-dimensional ECG phase vector through sinusoidal positional encoding. This is used for conditional control in subsequent networks. Its mapping relationship can be expressed as:

[0108] (1)

[0109] Here, MLP stands for Multilayer Perceptron, and Embedding represents temporal embedding operation.

[0110] In practice, the cardiac cycle position is read from the DICOM tag (0018, 9231). Assume certain data... (Mid-diastolic). The embedding operation uses sinusoidal position encoding and sets the dimensions. The MLP consists of two fully connected layers: the first layer maps 256 dimensions to 512 dimensions and then applies the ReLU activation function; the second layer maps the 512 dimensions back to 256 dimensions. The final output... It is a floating-point vector of length 256.

[0111] Step 2: Construct the Calcification Topology Prior.

[0112] This step aims to extract sparse calcification information from plain scan images to construct a "potential skeleton" of blood vessels.

[0113] Step 2.1, Calcification point extraction. Adaptive thresholding (e.g., >130 HU) is applied to the NCCT image to extract a set of high-density calcified pixels. The formula is as follows:

[0114] (2)

[0115] in, Represents the spatial domain of a three-dimensional image. Represents the coordinates in the image. Indicates the location of plain CT images grayscale value at that location The set calcification threshold.

[0116] In practice, a threshold is set. HU. To remove noise and non-vascular calcifications (such as bone) from images, connected component analysis is introduced after thresholding: the volume of each connected region is calculated, and regions smaller than 3 pixels (approximately) are removed. The tiny isolated points were identified, and the spine and rib regions were removed using a pre-generated lung mask, ultimately yielding a clean set of coronary artery calcification pixels. .

[0117] Step 2.2: Topological Path Generation. Using the Minimum Spanning Tree (MST) algorithm or Geodesic Distance Transform (GDT), discrete calcification clusters are connected to construct three-dimensional connected paths, generating a topological probability graph of the same size as the original image. The value represents the probability that the voxel belongs to the blood vessel centerline. The probability distribution function is defined as follows:

[0118] (3)

[0119] in The voxel coordinates in the image, To link the topological paths of calcification points, Voxel representation to path Euclidean distance, This is the attenuation coefficient. This step yields... It will be used as a strong anatomical constraint input to the neural network.

[0120] In practice, the Fast Marching Method is used to compute the set. The geodesic distances between points are used to construct a minimum spanning tree connecting the centers of all calcification clusters, thus forming a path. When generating the probability map, the attenuation coefficient α is set to 0.5. To reduce computational load, the cutoff distance is set to 5 voxels, i.e., when... At that time, directly ordered The final generated It is a three-dimensional probability matrix of size 128×128×64.

[0121] Step 3: Use Graph-Voxel Coupled Generative Network (GVC-Net) to achieve accurate CTA image generation;

[0122] Step 3.1: Divide the dataset into a training set, a validation set, and a test set;

[0123] Specifically, the 900 data points were randomly divided in an 8:1:1 ratio. During training, to adapt to memory limitations, a random cropping strategy was adopted, cropping the large input NCCT image into 128×128×64 three-dimensional image patches as network input.

[0124] Step 3.2: Construct a graph-voxel coupled generative network. The framework of the network is as follows: Figure 2 As shown, a dual-branch heterogeneous architecture is adopted, which specifically includes a voxel-Transformer global context branch and a topology-graph inference branch. The two interact dynamically through a bidirectional feature projection module.

[0125] Voxel-Transformer Branch: The network structure of this branch is as follows Figure 3 As shown, its configuration is designed to process dense 3D CT voxel data, aiming to capture the macroscopic geometry and long-range dependencies of the heart. Specifically, this branch uses a 3D Swin Transformer as the backbone network. After the raw NCCT image blocks enter the network, self-attention is calculated within non-overlapping local windows using a shifted window attention mechanism, establishing global contextual associations between voxels and generating multi-scale voxel feature maps containing rich anatomical semantics. Compared to traditional convolution, this mechanism can more effectively utilize information about the overall structure of the heart (such as the shape of the ventricles) to help infer the direction of blood vessels.

[0126] The specific implementation parameters are as follows: the Patch Partition layer divides the input into 4×4×4 blocks, and the feature dimension mapping is C=96. The network contains 4 stages, and the number of Swin Transformer Blocks is set to [2, 2, 6, 2]. The window size is set to 7×7×7. The number of heads for the multi-head self-attention mechanism is [3, 6, 12, 24]. The final output of this branch is... The feature map size is 32×32×16, and the number of channels is 768.

[0127] Topology-Graph Reasoning Branch: This branch is configured to handle sparse vascular structures in non-Euclidean space. The system treats the calcification points and centerline points constructed in step 2 as nodes V of a graph, and the topological connections between them as edges E, thus constructing a vascular topology graph. This branch employs a Graph Convolutional Network (GCN), enabling feature information to propagate and aggregate along the defined vascular topology. Its inter-layer propagation formula is expressed as:

[0128] (4)

[0129] in and Table 1 Layer input and output node feature matrices; Indicates the first A layer-trainable weight matrix is ​​used to perform linear transformations on the features; Represents a nonlinear activation function; This represents the adjacency matrix with added self-loops, where The original adjacency matrix describing the connectivity relationships between blood vessel nodes. It is the identity matrix; Then it is The degree matrix is ​​used to normalize features during convolution to maintain numerical stability. Through this graph-based feature propagation mechanism, even if a blood vessel is invisible in a plain scan image due to low density (i.e., lacking features in voxel space), its feature information can still be obtained through graph connectivity. Calcification nodes visible from upstream and downstream The information is transmitted over, thereby effectively repairing the broken blood vessels at the logical level and ensuring that the generated blood vessel tree has complete topological continuity.

[0130] The specific implementation parameters are as follows: N=1024 points are uniformly resampled along path L as graph nodes V. If the Euclidean distance between two nodes is less than 3mm, the corresponding position in the adjacency matrix A is set to 1. The GCN network contains 3 layers, and the output feature dimensions of each layer are [128, 256, 512]. Activation function Choose LeakyReLU (slope of negative half axis 0.2). The initial features are composed of the three-dimensional spatial coordinates (x, y, z) of the nodes.

[0131] Bidirectional Feature Projection Module: To achieve information complementarity between the two heterogeneous branches, the network is configured with a bidirectional projection mechanism. On one hand, it performs V2G (Voxel-to-Graph) projection: using a trilinear interpolation algorithm, based on the spatial coordinates of the graph nodes, it projects the Voxel feature map... The corresponding feature vectors are sampled and injected into the graph nodes, thus giving the abstract graph nodes specific texture context information. On the other hand, G2V (Graph-to-Voxel) projection is performed: using Gaussian Splatting technology, the graph node features after GCN inference are transferred to the graph nodes. Mapping back to three-dimensional voxel space creates a feature map with spatial indicative function, which is then superimposed onto the voxel branches. This is equivalent to lighting up a "vascular navigation light" in voxel space, forcing the network to focus on the anatomical region where the blood vessel is located.

[0132] In practical implementation, in V2G projection, for each graph node... According to its normalized coordinates in The mid-sampling yields a 768-dimensional vector, which is then reduced to 128 dimensions by a fully connected layer before being concatenated with node features. In G2V projection, the Gaussian kernel radius is set. A voxel scatters the features of 1024 nodes onto a 32×32×16 grid, and fills the feature values ​​of uncovered areas with 0.

[0133] ECG Modulation and Image Decoding: A hybrid feature map, fusing graph and voxel features, is then fed into the decoder. The decoder consists of cascaded residual convolutional modules, with an adaptive instance normalization (AdaIN) layer embedded in each module. This layer receives the ECG vector extracted in step 1. The mean and variance of the feature map are dynamically adjusted to ensure that the generated coronary artery morphology strictly corresponds to the cardiac cycle phase (such as diastole or systole) at the time of scanning.

[0134] In practice, the decoder contains four upsampling modules, each containing a 3×3×3 transpose convolution for upsampling, and two residual convolutional blocks. The AdaIN layer is located after each residual block, utilizing... The generated scaling factor and bias factor ,implement operate.

[0135] Step 3.3: Construct a joint frequency-spatial loss function to guide network training.

[0136] To address the problem of blurred edges in generated images, this invention defines a composite loss function. This function includes not only the L1 loss for constraining pixel consistency. and combat losses It also introduces frequency domain consistency loss. Graph topology loss Specifically, the frequency domain loss performs a 3D Fast Fourier Transform (FFT) on the generated and real images, constraining their distance in the amplitude spectrum to force the model to recover the high-frequency edge information of the blood vessel wall; the graph topology loss constrains the consistency between the generated blood vessel skeleton structure and the prior topology graph. The hybrid loss is defined as follows:

[0137] (5)

[0138] in, , , They represent , , In this embodiment, the weighting coefficients are... , and The values ​​are 10, 0.1, and 1. Additionally, Calculate the L2 norm distance between the amplitude spectrum of the generated image and the amplitude spectrum of the real image; Calculate the blood vessel segmentation results of the generated image and The Dice coefficient loss; Use least squares GAN (LSGAN) loss.

[0139] Step 3.4: Training the graph-voxel coupled network.

[0140] A phased training strategy is adopted: in the early stage of training, the voxel branch is frozen and the training graph branch is focused on learning the correct vascular topology; then all parameters are unfrozen and joint training is performed using the AdamW optimizer until the vascular structure similarity index (SSIM) on the validation set converges, and the final generative model is obtained.

[0141] In practice, an NVIDIA A100 GPU was used for training. The optimizer was AdamW, with the following parameters: , The weights decay to 1e−4. The initial learning rate is 2×10. −4 A cosine annealing strategy was employed. The first phase (first 50 epochs) optimized only the GCN branch parameters; the second phase (epochs 51-200) jointly optimized all parameters. The batch size was set to 4. Training was stopped and the model was saved when the SSIM metric on the validation set no longer improved for 10 consecutive epochs.

[0142] This embodiment provides a method for generating coronary CTA images using ECG-gated cardiac CT scans based on deep learning. A Graph-Voxel Coupled Generative Network (GVC-Net) is used to achieve accurate CTA image generation. A collaborative training strategy integrating voxel and graph branches is employed. The voxel branch uses a 3D Swin Transformer network to capture the macroscopic geometry and global contextual information of the heart. The graph branch uses a topological graph constructed based on calcification points as input, employing a Graph Convolutional Network (GCN) for explicit inference in non-Euclidean space and introducing a bidirectional feature projection module to achieve deep feature interaction between the voxel and topological domains, thereby enhancing the model's ability to repair small vessel ruptures. Finally, the features from both branches are fused and combined with ECG phase modulation for coordinated end-to-end training.

[0143] Example 2

[0144] This embodiment provides a system for generating coronary artery CTA images based on deep learning using ECG-gated cardiac CT scans. Specifically, taking coronary artery vascular tree reconstruction as an example, it includes: a multimodal data preprocessing module, a calcification topology prior construction module, and a graph-voxel coupling generation inference module.

[0145] The multimodal data preprocessing module specifically includes:

[0146] This module is configured to integrate a DICOM parsing unit and a 3D registration unit. The DICOM parsing unit, built on Python's Pydicom library, is configured to read the input data's tag (0018, 1088) to obtain heart rate information and read the tag (0018, 9231) to obtain cardiac cycle location information. The 3D registration unit, built on the SimpleITK open-source library, is configured to execute a multi-resolution B-spline freeform deformation (FFD) algorithm. Using mutual information as the optimization objective function, it performs sub-pixel-level 3D elastic registration between the input plain CT image and a standard anatomical template, outputting isotropic image data resampled to 0.5mm × 0.5mm × 0.5mm.

[0147] The aforementioned calcification topology prior construction module specifically includes:

[0148] This module is configured to integrate an adaptive thresholding segmentation algorithm unit and a graph theory path search algorithm unit. The adaptive thresholding segmentation algorithm unit is configured to automatically extract high-density pixels with CT values ​​greater than 130 HU and perform connected component volume filtering to remove noise points smaller than 3 mm. The graph theory path search algorithm unit is configured to run a fast traversal method to calculate the geodesic distance between the center points of calcification clusters and use the minimum spanning tree algorithm to construct a three-dimensional vascular skeleton. The final output is a coordinate matrix containing N nodes and an N×N adjacency matrix as topology input.

[0149] The graph-voxel coupling generation inference module specifically includes:

[0150] This module is built on the PyTorch deep learning inference framework and internally stores the weight files (.pth) of the converged GVC-Net model. At runtime, this module is configured to feed the input image patches and topology map into the SwinTransformer branch and GCN branch for parallel computation, respectively. The module embeds a bidirectional feature projection operator, configured to perform V2G (trilinear interpolation) and G2V (Gaussian sputtering) operations to achieve feature fusion. The decoder unit is configured to receive the ECG vector output from the preprocessing module and dynamically adjust the vascular morphology of the generated image through the AdaIN layer. This module is also configured to support CUDA acceleration, with a single inference time of approximately 1.5 seconds in an NVIDIA A100 GPU environment, and supports automatically stitching the output image patches into a complete 3D cardiac CTA image.

[0151] This invention is not limited to coronary angiogenesis, but can also be applied to medical image modality conversion tasks that rely on prior calcification or tubular structures, such as cerebral angiography and lower limb arterial angiogenesis.

[0152] Example 3

[0153] This embodiment provides an electronic device, including: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to realize the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling.

[0154] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0155] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0156] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0157] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0158] Example 4

[0159] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling.

[0160] For example, if the modules / units integrated in the electronic device described in Embodiment 3 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the method of Embodiment 1 by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted from the content as required by the legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium may not include electrical carrier signals and telecommunication signals.

[0161] In summary, the present invention can achieve the following effects: (1) Eliminate contrast agent-related risks: enable patients with renal insufficiency, contrast agent allergy, etc. to obtain effective coronary artery imaging assessment; (2) Reduce radiation dose: avoid a complete CTA scan and reduce the overall radiation dose to patients; (3) Improve examination efficiency: simplify the examination process and obtain dual information on calcium score and vascular morphology in a single plain scan; (4) Ensure diagnostic accuracy: the generated images are highly consistent with real CTA images in terms of vascular contrast, lumen display and lesion characteristics, and have clinical diagnostic value.

[0162] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling, characterized in that, Includes the following steps: S1. Construct a multimodal spatiotemporally aligned cardiac CT dataset: Collect ECG-gated plain CT images and coronary CTA images of the same patient, and perform three-dimensional elastic registration of the plain CT images based on the CTA images; extract ECG phase information from the header file of the original DICOM data and encode it into a high-dimensional ECG phase vector; S2. Constructing a calcification topology prior map: On the registered plain CT image, calcification point set is extracted by threshold segmentation; the calcification point set is connected based on the topology path generation algorithm to form a three-dimensional continuous path, and a topology probability map reflecting the probability distribution of the vessel centerline is calculated and generated. S3. Generate CTA images using a graph-voxel coupling network: Input the registered plain CT image and the topological probability map into a pre-trained graph-voxel coupling network, and inject the ECG phase vector for conditional control to generate the target coronary artery CTA image. The graph-voxel coupling generation network adopts a dual-branch heterogeneous architecture, including: The voxel-Transformer branch is used to extract global context voxel features from plain CT images; The topology-graph reasoning branch is used for feature reasoning and propagation on a vascular topology consisting of calcifications and centerline points; A bidirectional feature projection module is used to realize bidirectional feature mapping and fusion between the voxel-Transformer branch and the topology-graph reasoning branch; The conditional decoder is used to decode and output coronary CTA images based on fusion features and ECG phase vectors.

2. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, In step S1, the ECG phase information is encoded into a high-dimensional ECG phase vector, as follows: Extract the percentage of the ECG RR interval at the scan time from the header file of the original DICOM data. ; The percentage of the RR interval Through an embedding layer containing a multilayer perceptron, it is mapped to a high-dimensional ECG phase vector of a set length. The mapping relationship is expressed as: （1） in, This represents a multilayer perceptron; This indicates a phase embedding operation, using sinusoidal position coding.

3. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, Step S2 specifically includes: S21. Calcification point extraction: Apply adaptive threshold segmentation to plain CT images and extract calcification points with gray values ​​higher than a set threshold. The pixels constitute a three-dimensional calcification point set. : （2） in, Represents the spatial domain of a three-dimensional image. Represents the position coordinates in the image. Indicates the position coordinates of a plain CT image. The grayscale value at that location; S22. Topological Probability Graph Generation: Using the minimum spanning tree algorithm or the geodesic distance transformation algorithm, the set of calcified points is connected to generate a continuous three-dimensional topological path. ; Calculate the topological probability graph The formula is as follows: （3） in, In voxel coordinates, Voxel representation to path Euclidean distance, This is the attenuation coefficient.

4. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, In step S3, the voxel-Transformer branch of the graph-voxel coupling generation network is set to process voxel data from plain CT images to capture the geometric structure and long-range dependencies of the heart. Specifically: Using a 3D Swin Transformer as the backbone network, after plain CT images are input into this network, self-attention is calculated within non-overlapping local windows through a moving window attention mechanism to establish global contextual associations between voxels and generate multi-scale voxel feature maps containing anatomical semantics. .

5. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, In step S3, the topology-graph inference branch of the graph-voxel coupled network is configured to use a graph convolutional network to process the sparse topology graph and explicitly infer the connectivity of the vascular tree in non-Euclidean space, as detailed below: Extract calcification points and centerline points from the topological probability graph obtained in step 2. Use the extracted calcification points and centerline points as graph nodes V, and the topological connections between nodes as edges E to construct a vascular topological graph. ; Graph Convolutional Network (GCN) is used to propagate and aggregate node feature information along the blood vessel topology. The inter-layer propagation formula is as follows: （4） in, , Table 1 Layer input and output node feature matrices; Indicates the first The weight matrix of the layer training is used to perform linear transformations on the features; Represents a nonlinear activation function; This represents the adjacency matrix with added self-loops. The original adjacency matrix describing the connectivity relationships between blood vessel nodes. It is the identity matrix; Then it is The degree matrix is ​​used to normalize features during the convolution process.

6. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, In step S3, the bidirectional feature projection module of the graph-voxel coupling generation network performs bidirectional feature mapping to obtain a hybrid feature map, specifically including: Voxel to graph projection: Based on the 3D coordinates of the graph nodes, the corresponding feature values ​​are sampled from the voxel feature map output by the voxel-Transformer branch through trilinear interpolation and added to the graph node features; Graph-to-voxel projection: Projecting the graph node features output by the topology-graph inference branch. The feature map is rendered into a three-dimensional voxel space using Gaussian sputtering technology, and then fused with the feature map of the voxel-Transformer branch.

7. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, In step S3, the conditional decoder of the graph-voxel coupled network contains multiple residual convolutional modules, and an adaptive instance normalization layer is embedded in each residual convolutional module. The adaptive instance normalization layer dynamically adjusts the mean and variance of the mixed feature map based on the input ECG phase vector, so that the generated coronary artery morphology corresponds to the cardiac cycle phase at the scan time.

8. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 1, characterized in that, Before step S3, the method further includes training the graph-voxel coupled generation network, specifically: Construct a training set that includes plain CT images, topological probability maps, ECG phase vectors, and real CTA images; The graph-voxel coupled generation network is trained using a composite loss function. for: （5） in, Indicating resistance to loss, Indicates L1 loss, This represents the frequency domain consistency loss. Represents graph topology loss. , , They represent , , The weighting coefficients.

9. The method for generating coronary CTA images based on calcification topological prior and graph-voxel coupling according to claim 8, characterized in that, The training process of the graph-voxel coupled generation network adopts a phased strategy: In the first stage, the parameters of the voxel-Transformer branch are frozen, and only the topology-graph inference branch is trained to learn the correct vascular topology. In the second stage, all network parameters are unfrozen, and the composite loss function is used to jointly optimize and train the overall network until the similarity index of blood vessel topology on the validation set converges, thus obtaining the final graph-voxel coupled generative network.

10. A coronary artery CTA image generation system based on calcification topological prior and graph-voxel coupling, characterized in that, This system is used to implement the coronary CTA image generation method based on calcification topology prior and graph-voxel coupling as described in any one of claims 1 to 9. The system includes a multimodal data preprocessing module, a calcification topology prior construction module, and a graph-voxel coupling generation inference module, wherein: The multimodal data preprocessing module is configured to parse the acquired DICOM data, complete the three-dimensional elastic registration of the plain CT images, and extract the ECG phase vector. The calcification topology prior construction module is configured to automatically identify calcification regions in plain CT images and construct a three-dimensional vascular centerline path and corresponding adjacency matrix based on graph theory algorithms. The graph-voxel coupling generation inference module loads the trained graph-voxel coupling generation network model, runs the voxel-transformer branch to process voxel data and the topology-graph inference branch to process topology data in parallel, and realizes feature interaction through the bidirectional feature projection module. Finally, it directly outputs virtual coronary CTA images through the ECG-modulated conditional decoder.

11. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to implement the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling as described in any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the coronary CTA image generation method based on calcification topological prior and graph-voxel coupling as described in any one of claims 1 to 9.

Images