A CTA image generation method based on a state space diffusion network
By employing a state-space diffusion network-based approach, combined with an improved state-space model and discriminant network, the discontinuity problem in vascular structure reconstruction during CTA image generation was resolved. This approach enabled the generation of high-quality contrast-free CTA images, adapting to different devices and scanning conditions, and improving image fidelity and clinical readability.
Patent Information
- Application Number
- CN202610722101.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-25
- Publication Date
- 2026-08-25
AI Technical Summary
Existing CTA image generation methods struggle to effectively reconstruct vascular structures without contrast agents, exhibiting issues such as vascular rupture, topological errors, and anatomical inconsistencies. Furthermore, their high equipment dependence limits their application in resource-constrained scenarios.
A state-space diffusion network-based approach is adopted, which combines an improved state-space model and a discriminant network. Through adversarial mechanisms, joint optimization is performed to generate CTA images. A multi-loss function collaborative optimization strategy is used to enhance the continuity and topological consistency of vascular structures.
The generated CTA images significantly improve the accuracy of vascular structure reconstruction, image fidelity, and clinical readability. They can accurately restore vascular structure information without contrast agents, adapt to different equipment and scanning conditions, and have good stability and generalization ability.
Smart Images
Figure CN122636802A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image processing, and in particular to a method for generating CTA images based on state-space diffusion networks. Background Technology
[0002] Imaging assessment of acute ischemic stroke (AIS) is highly dependent on clear representation of vascular structures, especially rapid identification of large vessel occlusion (LVO) and assessment of collateral circulation status.
[0003] Computed Tomography Angiography (CTA) is the most commonly used vascular imaging technique in clinical practice, providing high-contrast visualization of cerebral blood vessels. However, CTA requires the injection of exogenous iodine contrast agents, which not only increases the complexity of the imaging process but also generates additional radiation and limits its use in certain special populations. Furthermore, CTA has high requirements for equipment and operation, and its accessibility is limited in resource-constrained scenarios.
[0004] Non-contrast CT (NCCT), as a widely available basic imaging method, has advantages such as being non-invasive, fast, and easy to operate. However, due to the extremely low natural contrast between blood and the surrounding brain parenchyma, NCCT struggles to directly display vascular morphology. Therefore, how to reconstruct vascular images equivalent to CTA from NCCT without contrast agents has become an important research direction in the field of medical image computing. In recent years, deep learning technology has made significant progress in cross-modal medical image generation tasks, with the following being the mainstream methods: (1) Convolutional Neural Network (CNN) and U-Net structure: Image reconstruction is achieved through local convolution and encoder-decoder structure, but due to the limited receptive field, it is difficult to effectively capture the structural dependencies between distant branches in the vascular tree, resulting in the easy loss or discontinuity of distal blood vessels.
[0005] (2) Generative Adversarial Networks (GANs): such as Pix2pix and CycleGAN, improve the realism of the overall image distribution through adversarial training between the generator and the discriminator. However, this method is prone to vascular rupture, topological errors, or phantom blood vessels, making it difficult to guarantee the continuity and anatomical consistency of the vascular structure in the generated image. In addition, this method has the problem of training instability, which can easily lead to non-physiological textures or overly sharp edges, reducing the visual credibility of the generated image. Summary of the Invention
[0006] To address the aforementioned problems and technical requirements, this application proposes a CTA image generation method based on a state-space diffusion network. The technical solution of this application is as follows: A CTA image generation method based on state-space diffusion networks, the CTA image generation method comprising: Acquire multiple sample image pairs, each sample image pair including a paired sample non-enhanced CT image and a sample real CTA image; Random noise is added to the sample non-enhanced CT image and then input into the state space diffusion network for several steps of denoising to obtain the sample reconstructed CTA image. The state space diffusion network adopts the Unet structure and the Unet structure uses an improved state space model for image feature extraction. The real CTA image and the reconstructed CTA image of the sample are input into the discriminant network for block discrimination. The discriminant network and the state space diffusion network are jointly optimized through an adversarial mechanism. The unenhanced CT image of the patient's target area is input into the trained state-space diffusion network to generate a CTA image of the patient's target area.
[0007] A further technical solution is that, after the improved state-space model in the state-space diffusion network divides the input feature map into multiple image blocks, the image blocks are reordered according to the vascular region response feature map of the real CTA image of the sample and the vascular structure. The image features of each reordered image block are mapped into high-dimensional features, and the output feature map is obtained by using the continuous-time state-space model to model the high-dimensional feature sequence.
[0008] A further technical solution involves using a continuous-time state-space model to model the image features of a high-dimensional feature sequence, resulting in an output feature map including: Based on the high-dimensional feature sequence, the first... Image features of image blocks Dynamically generate the parameter matrix in the continuous-time state-space model to obtain the result with respect to the first... A continuous-time state-space model for matching image patches, comprising a state evolution equation and an output equation; Using a zero-order hold strategy for the first The continuous-time state-space model of the nth image patch matching is discretized to obtain the discrete state update equation, which is then used to... The discrete state update equation for matching the nth image patch will be the nth High-dimensional features of image patches Encoded as a hidden state vector The output signal is obtained after decoding. ; The output feature map is obtained by remapping the output signals of each image block in the high-dimensional feature sequence.
[0009] Its further technical solution is, the first The discrete state update equations for matching image patches include:
[0010] in, It is a parameter matrix Discretized form, It is a parameter matrix Discretized form, and These are parameter matrices, It is based on the first Image features of image blocks Dynamically generated continuous weights It is the first The hidden state vector obtained by encoding the previous image block of each image block. It is the symbol for element-wise multiplication.
[0011] Its further technical solution is,
[0012]
[0013] in, It is based on the first Image features of image blocks Dynamically generated time step It is an identity matrix.
[0014] Its further technical solution involves joint optimization using a discriminative network and a state-space diffusion network through an adversarial mechanism, including: Reconstructing vascular region response feature maps from CTA images based on samples Vascular region response feature map of real CTA images of the sample Calculate vascular remodeling loss; The total loss function is obtained by weighting the vascular remodeling loss, adversarial loss, and cycle consistency loss. The state space diffusion network and the discriminant network are then jointly optimized according to the total loss function.
[0015] A further technical solution involves calculating vascular remodeling loss, including: Calculate the loss of vascular structure sensing , Indicates calculation L1 loss, It is a response weight map of vascular regions, and the response weight of vascular regions is higher than that of non-vascular regions. It is the symbol for element-wise multiplication.
[0016] A further technical solution involves calculating vascular remodeling loss, including: Calculate the enhancement loss at the vessel edge , It is a vascular region response feature map gradient plot, It is a vascular region response feature map gradient plot, Indicates calculation L1 loss.
[0017] A further technical solution involves calculating vascular remodeling loss, including: Calculate the loss of vascular continuity constraint , Indicates calculation L1 loss.
[0018] A further technical solution involves acquiring multiple sample image pairs, including: Multiple initial image pairs were obtained by scanning with different scanning devices under different scanning protocols and conditions. Each initial image pair included a non-contrast CT image of the same patient site and a real CTA image of the same patient site, and the data was preprocessed. After grayscale normalization of the images in each initial image pair, the three-dimensional image registration method is used to spatially align the sample non-enhanced CT image and the sample real CTA image in the same initial image pair. The initial image pairs that have completed image registration are then resampled and size-unified to obtain sample image pairs. Data augmentation is performed on the processed sample image pairs.
[0019] The beneficial technical effects of this application are: This application discloses a CTA image generation method based on a state-space diffusion network. This method uses a state-space diffusion network to generate CTA images from unenhanced CT images. The state-space diffusion network introduces an improved state-space model during the diffusion generation process to enhance the modeling ability of long-range dependent information, thereby improving the continuity and topological consistency of vascular structures. At the same time, adversarial learning mechanisms are combined in the generation framework to enhance the image detail expression ability through local region discrimination strategies, making the generated CTA images closer to real CTA images in terms of boundary clarity and texture restoration. Compared with existing technical solutions based on convolutional neural networks or generative adversarial networks, the CTA images generated using the state-space diffusion network of this application have significantly improved vascular structure reconstruction accuracy, image fidelity, and clinical readability. This enables high-quality conversion from unenhanced CT images to CTA images, accurately restoring vascular structure information without the need for contrast agents, and has good technical advancement and clinical application value.
[0020] This method optimizes the state-space model introduced during the diffusion generation process based on the traditional architecture. It employs a structure-aware feature sequence generation mechanism, making the generated sequences spatially more consistent with the continuous distribution of potential vascular structures, thus providing structural priors for subsequent modeling. Furthermore, it introduces a feature-adaptive state parameter modeling mechanism and a spatially adaptive step size parameter, enabling the model to adaptively adjust the state evolution process based on the feature differences between vascular and non-vascular regions, thereby enhancing its ability to model slender structures. The introduction of a structural continuity enhancement term during state updates further strengthens the topological consistency of the vascular structure, enhancing the dependencies between states within the vascular region, effectively maintaining the continuity and branching characteristics of the blood vessels. This improved state-space model effectively reduces common problems in traditional generation methods such as vascular breakage, artifacts, and structural mismatches.
[0021] This method also employs a multi-loss function collaborative optimization strategy, using not only cyclic consistency loss and adversarial loss, but also vascular structure perception loss, vascular edge enhancement loss, and vascular continuity constraint loss. While ensuring overall structural consistency, it also considers local detail restoration, effectively solving problems such as structural distortion, missing small vascular branches, and artifacts present in traditional generation methods. This results in generated CTA images that are highly consistent with real CTA in terms of overall grayscale distribution, local contrast, and vascular boundary clarity. In particular, it demonstrates excellent detail restoration capabilities in key anatomical regions (such as the internal carotid artery bifurcation and the vertebrobasilar artery system) and the presentation of small vascular branches, thus significantly improving image fidelity.
[0022] This method exhibits good stability and generalization ability in multi-center data environments and can adapt to different devices and scanning conditions. Attached Figure Description
[0023] Figure 1 This is a flowchart of the training phase of a state-space diffusion network in one embodiment of this application.
[0024] Figure 2 This is a comparison of CTA images generated from unenhanced CT images and real CTA images using the state space diffusion network, Pix2pix model, and RegGAN model of this application in one example of this application. Detailed Implementation
[0025] The specific embodiments of this application will be further described below with reference to the accompanying drawings.
[0026] This application discloses a CTA image generation method based on a state-space diffusion network. This method utilizes a state-space diffusion network to automatically generate corresponding CTA images from non-contrast CT images, thereby obtaining clear vascular structure information without the need for contrast agents. The method first requires training the state-space diffusion network. The training phase of the state-space diffusion network includes the following steps, please refer to [reference needed]. Figure 1 The flowchart shown is as follows: Step 1: Obtain multiple sample image pairs. Each sample image pair includes a paired sample non-enhanced CT image and a sample real CTA image.
[0027] In practical applications, initial image pairs are first obtained by scanning the same patient site using a scanning device, producing a non-contrast CT image and a true CTA image. These initial image pairs are then preprocessed to remove images with obvious artifacts or substandard quality. Furthermore, to improve the adaptability of the trained state-space diffusion network, initial image pairs are obtained by scanning with different devices under different scanning protocols and conditions, including conditions specific to different patient populations.
[0028] After grayscale normalization of the images in each initial image pair, a 3D image registration method is used to spatially align the sample non-enhanced CT image and the sample real CTA image within the same initial image pair. This ensures a strict spatial correspondence between the sample non-enhanced CT image and the sample real CTA image in the same coordinate system, guaranteeing pixel-level consistency during subsequent model learning. Then, the initial image pairs that have completed image registration are resampled and resized to obtain sample image pairs that meet the input requirements of subsequent models.
[0029] The processed sample image pairs are augmented with data to improve the model's generalization ability. The data augmentation operations performed include rotation, translation, and mirroring.
[0030] Step 2: After adding random noise to the sample non-enhanced CT image, input it into the state space diffusion network for several denoising steps to obtain the sample reconstructed CTA image.
[0031] This application uses a diffusion model framework to add random noise to sample non-enhanced CT images, and then uses a state-space diffusion network for reverse denoising to generate CTA images. The forward process gradually transforms the input sample non-enhanced CT image into Gaussian noise, while the reverse process gradually recovers and generates the CTA image through the state-space diffusion network. This process is characterized by high stability and high generation quality. The specific operation of adding random noise to the sample non-enhanced CT image can be referred to the processing methods of existing diffusion models, and will not be elaborated here.
[0032] In practical applications, state-space diffusion networks are used to perform multiple denoising processes on unenhanced CT images of samples with added random noise, such as... Figure 1 As shown, random noise is added to the sample non-enhanced CT image to obtain a feature map, which is then input into the state space diffusion network. After a certain number of denoising steps, the feature map is obtained again. This feature map is then input into the state space diffusion network again, and after a certain number of denoising steps, it is input into the state space diffusion network again, and so on. The final output feature map is the sample reconstructed CTA image. Figure 1 Take the two-stage denoising process using a state-space diffusion network as an example.
[0033] The State Space Diffusion Network (USDN) employs the UNet architecture, which uses an encoder-decoder network structure and combines residual connections and skip connections to enhance feature representation capabilities. The encoder consists of multiple convolutional layers, with multiple residual modules introduced in the intermediate layers to improve the network's deep representation capabilities and enhance training stability. The decoder gradually restores image resolution through deconvolution operations and combines skip connections to fuse shallow and deep features, thereby preserving detailed information while maintaining global structural consistency.
[0034] To enhance the representation of vascular structures by state-space diffusion networks, this application optimizes the classic Unet structure by implementing the feature extraction unit using an improved state-space model. This improved state-space model is inserted into the key feature layers of the encoder and decoder in the Unet structure, enabling joint modeling of local details and global structure. Specifically, the improved state-space model replaces the attention module in the traditional Unet structure to capture spatial context information, enhancing the modeling and reconstruction capabilities for small blood vessels and weak edge structures. The specific implementation methods of the encoder-decoder structure, residual connections, and skip connections in the Unet structure can be adjusted according to application requirements. However, regardless of the structure used, it should possess multi-scale feature fusion and fine-grained structure recovery capabilities to ensure the integrity and continuity of small structures such as blood vessels during the generation process.
[0035] This application uses an improved state-space model for image feature modeling, which includes the following improvements compared to the classic state-space model: Unlike traditional methods that simply partition the input feature map into a sequence, the improved state-space model in the state-space diffusion network divides the input feature map into multiple image blocks. Then, based on the vascular region response feature map of the real CTA images, these image blocks are reordered according to vascular structure. This results in a sequence that spatially better reflects the continuous distribution of vascular structure, providing structural priors for subsequent modeling. Building upon this, the image features of each reordered image block are mapped to high-dimensional features. Finally, a continuous-time state-space model is used to model the high-dimensional feature sequence to obtain the output feature map.
[0036] The classic state-space model uses a continuous-time state-space model with a fixed parameter matrix. This model includes state evolution equations and output equations, expressed as follows:
[0037] in, It is the first High-dimensional features of image patches , For high-dimensional features The encoded hidden state vector, It is the hidden state vector Time derivative, For the hidden state vector The output signal obtained by decoding . , , , It is a learnable parameter matrix in a continuous-time state-space model. Traditional state-space models use a fixed parameter matrix. , , , To enable the model to have differentiated dynamic modeling capabilities in vascular and non-vascular regions, thereby enhancing its ability to represent slender structures, a spatial adaptive state parameter modulation mechanism is designed in this embodiment. When using a continuous-time state-space model to model image features of high-dimensional feature sequences, the mechanism modulates the state parameter based on the first parameter in the high-dimensional feature sequence. Image features of image blocks Dynamically generate the parameter matrix in the continuous-time state-space model, thereby obtaining the result related to the first... The continuous-time state-space model for matching image patches is written as:
[0038] in, , , , It is related to image features The parameter matrix for dynamic adaptive matching.
[0039] Considering that medical images are discrete sampled signals, this embodiment employs a zero-order hold strategy for the first... The continuous-time state-space model of the nth image patch matching is discretized to obtain the discrete state update equation, and then the nth image patch matching is used to obtain the discrete state update equation. The discrete state update equation for matching the nth image patch will be the nth High-dimensional features of image patches Encoded as a hidden state vector The output signal is obtained after decoding. Finally, the output signals of each image block in the high-dimensional feature sequence are remapped to obtain the output feature map.
[0040] The first one obtained after discretization The discrete state update equation for each image patch is written as:
[0041] in, It is a parameter matrix Discretized form, It is a parameter matrix Discretized form, and These are parameter matrices. It is the first The hidden state vector is obtained by encoding the previous image block of each image block.
[0042] In another embodiment, a spatially adaptive step size parameter is further introduced, that is, according to the first... Image features of image blocks Dynamically generate time step and according to time step For the The continuous-time state-space model for matching image patches is discretized, and the parameter matrix in the discretized form of the above discrete state update equation is then obtained. and Written as:
[0043] in, It is an identity matrix.
[0044] To further enhance the topological consistency of the vascular structure, in another embodiment, a structural continuity enhancement term is introduced into the state evolution equation. In this embodiment, the discretized result is... The discrete state update equation for each image patch is:
[0045] in, It is based on the first Image features of image blocks Dynamically generated continuous weights It is the symbol for element-wise multiplication. It is a structural continuity enhancement term added to the state evolution equation. By adding this structural continuity enhancement term, the information transmission intensity between adjacent states can be adjusted. This mechanism can enhance the dependency between states in the vascular region, thereby effectively maintaining the continuity and branching structure characteristics of the blood vessels.
[0046] Step 3: Input the real CTA image of the sample and the reconstructed CTA image of the sample into the discriminant network for block discrimination, and use the discriminant network and the state space diffusion network for joint optimization through an adversarial mechanism.
[0047] To improve the realism of the generated CTA images, this application introduces a discriminative network and adopts a local region-based discrimination strategy. This discriminative network, by segmenting the input real CTA image and the reconstructed CTA image into blocks, can more sensitively capture blood vessel boundaries and detailed textures, thereby prompting the state space diffusion network to output more realistic, continuous, and structurally sound reconstructed CTA images. During training, the state space diffusion network and the discriminative network jointly optimize through an adversarial mechanism, enabling the generated reconstructed CTA images to gradually approximate the real CTA images.
[0048] Specifically, this application improves the structural accuracy and clinical usability of the generated CTA images by jointly optimizing the model using multiple specialized loss functions oriented towards vascular structural features, including: First, calculate the cycle consistency loss; the specific calculation formula can be found in classical calculation methods. The cycle consistency loss is used to constrain the consistency between the mapping from non-enhanced CT images to CTA images and its inverse mapping, thereby ensuring that the overall structural information does not shift.
[0049] Next, the adversarial loss is calculated; the specific calculation formula can be found in classical methods. The adversarial loss is used to improve the realism and distribution consistency of the generated image.
[0050] In addition, to further enhance the expressive power of vascular structures, vascular region response feature maps of CTA images were reconstructed based on the samples. Vascular region response feature map of real CTA images of the sample The vascular remodeling loss is calculated, including: (1) Calculate the loss of vascular structure perception :
[0051] in, Indicates calculation L1 loss. It is a response weight map of vascular regions, and the response weight of vascular regions is higher than that of non-vascular regions. It is an element-wise multiplication symbol. (Based on the vascular region response weight map) By assigning higher weights to vascular regions, the model is guided to focus on optimizing vascular structures.
[0052] (2) Calculate the enhancement loss at the vessel edge :
[0053] in, Indicates calculation L1 loss. It is a vascular region response feature map gradient plot, It is a vascular region response feature map The gradient plot.
[0054] The loss of enhancement at the vascular margin Constraining the gradient information of the reconstructed CTA image and the real CTA image of the sample is beneficial to improving the clarity of the blood vessel boundary.
[0055] (3) Calculate the loss of vascular continuity constraint :
[0056] in, Indicates calculation L1 loss. This is a loss of vascular continuity constraint. It can enhance the structural consistency between adjacent regions, thereby helping to maintain the connectivity of vascular structures.
[0057] Finally, the total loss function is obtained by weighting the various vascular remodeling losses, adversarial losses, and cycle consistency losses, and the state space diffusion network and the discriminant network are jointly optimized according to the total loss function.
[0058] After completing the network training according to the above method, during the model inference stage, the non-enhanced CT image of the patient's target area is input into the trained state space diffusion network. The CTA image of the patient's target area can be gradually generated through the diffusion inverse denoising process. The generated CTA image can clearly show the course and branching structure of intracranial and extracranial blood vessels and has good clinical readability.
[0059] To demonstrate the validity of this application, in one example, unenhanced CT images (NCCT) were input into the state-space diffusion network, Pix2pix model, and RegGAN model of this application to generate CTA images, respectively. Real CTA images were then used as the gold standard reference images for comparison. The comparison results of three different cases or different anatomical levels are shown below. Figure 2 As shown. The red dashed boxes in each group of images mark the regions of interest (ROIs). The images below are magnified views of the corresponding ROIs. Red arrows indicate the locations of key vascular structures or detailed areas, allowing observation of differences in vascular visualization, boundary clarity, continuity, and the reconstruction of fine branches using different methods. Figure 2 The results show that the original unenhanced CT images have low contrast between blood vessels and surrounding soft tissue, making it difficult to clearly identify the target vascular structure. While the CTA images generated by the Pix2pix and RegGAN models enhance vascular visualization to some extent, they still suffer from problems such as blurred vessel boundaries, distortion of local structures, unclear display of small vessel branches, or insufficient continuity. In contrast, the CTA images generated by the state-space diffusion network in this application are closer to real CTA images in terms of vascular enhancement, edge sharpness, restoration of local details, and continuity of vessel course. They can more accurately restore the morphological features of the target blood vessels, demonstrating higher image fidelity and better clinical readability.
[0060] It should be noted that this application utilizes the noise addition and denoising framework in the diffusion model to progressively model and realize the gradual reconstruction process of vascular structures. This progressive reconstruction mechanism can also be implemented by other generation methods, such as generative models based on latent variable progressive optimization or multi-stage reconstruction strategies. It should be noted that regardless of the generation mechanism used, it should possess the ability to progressively refine and reconstruct vascular structures and work synergistically with the structural constraint mechanism of this application. Regarding feature modeling, this application employs an improved state-space model to enhance the expressive power of vascular structures through a spatially adaptive dynamic parameter modeling mechanism. This modeling method can also be replaced by other structures capable of achieving location-related parameter modulation or global dependency modeling, such as adaptive weighting methods based on attention mechanisms or sequence modeling methods. However, the above alternatives must still satisfy the requirement of being able to dynamically adjust the information propagation path according to the input features, thereby achieving effective modeling of vascular continuity and topology.
[0061] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.
Claims
1. A CTA image generation method based on state-space diffusion networks, characterized in that, The CTA image generation method includes: Acquire multiple sample image pairs, each sample image pair including a paired sample non-enhanced CT image and a sample real CTA image; Random noise is added to the sample non-enhanced CT image and then input into the state space diffusion network for several steps of denoising to obtain the sample reconstructed CTA image. The state space diffusion network adopts the Unet structure and the Unet structure uses an improved state space model for image feature extraction. The real CTA image and the reconstructed CTA image of the sample are input into the discriminant network for block discrimination. The discriminant network and the state space diffusion network are jointly optimized through an adversarial mechanism. The unenhanced CT image of the patient's target area is input into the trained state-space diffusion network to generate a CTA image of the patient's target area.
2. The CTA image generation method according to claim 1, characterized in that, The improved state-space model in the state-space diffusion network divides the input feature map into multiple image blocks. Based on the vascular region response feature map of the real CTA image, the image blocks are reordered according to the vascular structure. The image features of the reordered image blocks are mapped to high-dimensional features. The output feature map is obtained by modeling the high-dimensional feature sequence using the continuous-time state-space model.
3. The CTA image generation method according to claim 2, characterized in that, After using a continuous-time state-space model to model the image features of a high-dimensional feature sequence, the output feature map includes: Based on the high-dimensional feature sequence, the first... Image features of image blocks Dynamically generate the parameter matrix in the continuous-time state-space model to obtain the result with respect to the first... A continuous-time state-space model for matching image patches, comprising a state evolution equation and an output equation; Using a zero-order hold strategy for the first The continuous-time state-space model of the nth image patch matching is discretized to obtain the discrete state update equation, which is then used to... The discrete state update equation for matching the nth image patch will be the nth High-dimensional features of image patches Encoded as a hidden state vector The output signal is obtained after decoding. ; The output feature map is obtained by remapping the output signals of each image block in the high-dimensional feature sequence.
4. The CTA image generation method according to claim 3, characterized in that, No. The discrete state update equations for matching image patches include: in, It is a parameter matrix Discretized form, It is a parameter matrix Discretized form, and These are parameter matrices, It is based on the first Image features of image blocks Dynamically generated continuous weights It is the first The hidden state vector obtained by encoding the previous image block of each image block. It is the symbol for element-wise multiplication.
5. The CTA image generation method according to claim 4, characterized in that, in, It is based on the first Image features of image blocks Dynamically generated time step It is an identity matrix.
6. The CTA image generation method according to claim 3, characterized in that, Joint optimization using discriminative networks and state-space diffusion networks through adversarial mechanisms includes: Reconstructing vascular region response feature maps from CTA images based on samples Vascular region response feature map of real CTA images of the sample Calculate vascular remodeling loss; The total loss function is obtained by weighting the vascular remodeling loss, adversarial loss, and cycle consistency loss. The state space diffusion network and the discriminant network are then jointly optimized according to the total loss function.
7. The CTA image generation method according to claim 6, characterized in that, Calculating vascular remodeling loss includes: Calculate the loss of vascular structure sensing , Indicates calculation L1 loss, It is a response weight map of vascular regions, and the response weight of vascular regions is higher than that of non-vascular regions. It is the symbol for element-wise multiplication.
8. The CTA image generation method according to claim 6, characterized in that, Calculating vascular remodeling loss includes: Calculate the enhancement loss at the vessel edge , It is a vascular region response feature map gradient plot, It is a vascular region response feature map gradient plot, Indicates calculation L1 loss.
9. The CTA image generation method according to claim 6, characterized in that, Calculating vascular remodeling loss includes: Calculate the loss of vascular continuity constraint , Indicates calculation L1 loss.
10. The CTA image generation method according to claim 1, characterized in that, Obtaining multiple sample image pairs includes: Multiple initial image pairs were obtained by scanning with different scanning devices under different scanning protocols and conditions. Each initial image pair included a non-contrast CT image of the same patient site and a real CTA image of the same patient site, and the data was preprocessed. After grayscale normalization of the images in each initial image pair, the three-dimensional image registration method is used to spatially align the sample non-enhanced CT image and the sample real CTA image in the same initial image pair. The initial image pairs that have completed image registration are then resampled and size-unified to obtain sample image pairs. Data augmentation is performed on the processed sample image pairs.