Contrast-agent-free angiography generation system and method based on mask guidance

Abstract

The invention relates to a contrast-agent-free angiography generation system and method based on mask guidance, a parameter optimization device is used for performing parameter optimization on a conversion device by utilizing a loss function capable of highlighting a contrast agent shadow, and the capability of capturing complex image details of the conversion device is improved. Then, the Gaussian noise image is fused into a contrast-agent-free computed tomography image through a conversion device, computed tomography angiography is obtained through a fusion mode of feature sampling and gradual denoising, and it is guaranteed that the obtained computed tomography angiography has enough consistency and sense of reality; and the sharpness and fidelity of the generated vascular structure are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular, to a mask-guided contrast agent-free angiography generation system and method. Background Art

[0002] Computed tomography angiography (CTA) constructed based on the use of iodinated contrast agents (ICAs) can provide high-resolution and three-dimensional vascular images. Compared with non-contrast computed tomography images (NCCT), the greatest advantage is that computed tomography angiography can provide clearer and more detailed vascular anatomical images, thereby enhancing the contrast between blood vessels and surrounding tissues.

[0003] In order to convert non-contrast computed tomography images into computed tomography angiography, the prior art discloses a RegGAN model architecture, which enhances image generation quality, improves structural consistency, and reduces artifacts by introducing regularization techniques.

[0004] However, this model architecture is difficult to generate high-quality images with sufficient consistency and realism in complex imaging scenarios. Specifically, the computed tomography angiography generated by it has artifacts, which are manifested as irregular blood vessel shapes, inaccurate boundaries, and unnatural textures, thus weakening the clinical usability of the generated images. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to overcome the technical defect that the existing RegGAN model architecture is difficult to generate high-quality images with sufficient consistency and realism. To overcome this technical defect, the present invention provides a mask-guided contrast agent-free angiography generation system and method, specifically including a mask-guided contrast agent-free angiography generation system and a mask-guided contrast agent-free angiography generation method.

[0006] A mask-guided contrast agent-free angiography generation system provided by the present invention includes: A noise generation device configured to generate a Gaussian noise image by a random Gaussian noise iteration algorithm; An acquisition device configured to acquire non-contrast computed tomography images; A conversion device electrically connected to both the noise generation device and the acquisition device, configured to integrate the Gaussian noise image into the non-contrast computed tomography image and obtain computed tomography angiography through a fusion method of feature sampling and step-by-step denoising; A parameter optimization device, electrically connected to the conversion device, is configured to generate a binary contrast-enhanced mask based on a subtraction strategy using a data set, substitute the binary contrast-enhanced mask into a loss function for highlighting blood vessel regions affected by a contrast agent, and use the loss function to optimize the parameters of the conversion device, obtain model parameters, and transmit them to the conversion device.

[0007] The mask-guided contrast-agent-free angiography generation system disclosed in the present invention includes a noise generation device, an acquisition device, a conversion device, and a parameter optimization device. The parameter optimization device uses a loss function that can highlight the regions affected by the contrast agent to optimize the parameters of the conversion device, improving the ability of the conversion device to capture complex image details. The conversion device integrates a Gaussian noise image into a contrast-agent-free computed tomography (CT) image and obtains a CT angiography through a fusion method of feature sampling and step-by-step denoising, ensuring that the obtained CT angiography has sufficient consistency and realism. Incorporating the loss function that can highlight the regions affected by the contrast agent into the diffusion process based on a Gaussian noise image generated by a random Gaussian noise iteration algorithm not only effectively converts a contrast-agent-free CT image into a higher-quality CT angiography but also stabilizes the parameter optimization process, improves the sharpness and fidelity of the generated vascular structure, and thus overcomes the technical defects of the existing RegGAN model architecture.

[0008] In a possible implementation, the iteration formula of the random Gaussian noise iteration algorithm is as follows: , , , wherein, is a composite number representing the initially determined number of diffusion steps; is a composite primitive factor; represents the noise at the th iteration; represents the noise addition term at the th iteration; represents an arbitrarily small number; represents the conditional probability distribution of obtaining noise given noise ; represents the identity matrix; represents a Gaussian distribution; represents the exponential progress minimum value; represents the exponential progress maximum value.

[0009] When executing the above iterative formula, as the time step increases, the diffusion step will be much larger than 1, thereby reducing the number of diffusion steps. Eventually, the number of diffusion steps is reduced from T to T / k, greatly improving the generation efficiency of noise.

[0010] In a possible implementation manner, the conversion device is a communication network structure formed by connecting multiple adversarial diffusion generation modules in series. The adversarial diffusion generation module is configured to convert the input image into an output image through feature sampling by encoding and decoding and denoising by Bayesian rule sampling; furthermore, it can not only integrate Gaussian noise images into non-contrast computed tomography images, but also further ensure that the obtained computed tomography angiography has sufficient consistency and realism.

[0011] In a possible implementation manner, the adversarial diffusion generation module includes an encoder, a decoder, a Bayesian posterior sampling sub-module, a contrast enhancement attention guidance sub-module, and a contrast agent perception discrimination sub-module; The input end of the encoder in the adversarial diffusion generation module at the first position is electrically connected to the output end of the noise generation device and the output end of the acquisition device at the same time. The input end of the encoder in the remaining adversarial diffusion generation modules is electrically connected to the output end of the Bayesian posterior sampling sub-module in the previous adversarial diffusion generation module; In each adversarial diffusion generation module, the input end of the decoder is electrically connected to the output end of the encoder, the input end of the Bayesian posterior sampling sub-module is electrically connected to the output end of the decoder, the input end of the contrast agent perception discrimination sub-module is electrically connected to the output end of the Bayesian posterior sampling sub-module and the output end of the parameter optimization device at the same time, the input end of the contrast enhancement attention guidance sub-module is electrically connected to the output end of the decoder and the output end of the parameter optimization device at the same time, and the input end of the encoder is also electrically connected to the output end of the contrast agent perception discrimination sub-module and the output end of the contrast enhancement attention guidance sub-module at the same time.

[0012] In a possible implementation manner, the encoder is configured to perform a convolution operation on its input image to extract anatomical structure and texture information, and integrate time information during the extraction process, and output an encoded image; The decoder is configured to perform deconvolution denoising on the encoded image to obtain an estimated denoised image; The Bayesian posterior sampling sub-module is configured to denoise the estimated denoised image through Bayesian sampling rules to obtain a denoising result; The contrast enhancement attention guidance sub-module is configured to, after the parameter optimization device is started, multiply the estimated denoised image and the binary contrast enhancement mask element by element and then perform ReLU activation to generate an attention map, and optimize the parameters of the encoder and decoder in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum value of the distance between the attention map and the binary contrast enhancement mask; The angiography-aware discrimination sub-module is configured to, after the parameter optimization device is started, optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum value of a non-saturating adversarial loss function with gradient penalty.

[0013] The adversarial diffusion generation module with the above structure and functions can integrate a Gaussian noise image into a non-contrast computed tomography image through an encoder and a decoder, and perform denoising through the Bayesian posterior sampling sub-module. Combined with the included contrast enhancement attention guidance sub-module and angiography-aware discrimination sub-module, it can perform independent parameter optimization of the encoder, decoder, and Bayesian posterior sampling sub-module of each module, further ensuring that the obtained computed tomography angiography has sufficient consistency and realism.

[0014] In a possible implementation manner, the parameter optimization device includes: An unsupervised mask generation module, whose output end is electrically connected to the input ends of all the contrast enhancement attention guidance sub-modules and the input end of the angiography-aware discrimination sub-module at the same time, and is configured to generate a binary contrast enhancement mask based on a subtraction strategy using the data set; A mask-guided feature alignment module, whose input end is electrically connected to the output end of the unsupervised mask generation module and the output end of the Bayesian posterior sampling sub-module in the adversarial diffusion generation module at the end at the same time, and is configured to substitute the binary contrast enhancement mask into the loss function to optimize the parameters of the conversion device using the loss function, obtain model parameters, and transmit them to all the encoders, the decoders, and the Bayesian posterior sampling sub-modules.

[0015] In the parameter optimization device with the above structure, the supervised mask generation module uses the data set to generate a binary contrast enhancement mask, and the mask-guided feature alignment module realizes model parameter optimization, stabilizing the parameter optimization process and further improving the sharpness and fidelity of the generated vascular structure.

[0016] In a possible implementation manner, the calculation formula of the subtraction strategy is: , In the formula, represents the binary contrast enhancement mask of the th data in the said dataset; represents the computed tomography angiography of the th data in the said dataset; represents the non-contrast computed tomography image of the th data in the said dataset; represents binarization processing.

[0017] The binary contrast enhancement mask constructed by the above formula is actually an unsupervised contrast enhancement mask. Combining the introduced contrast perception technology, it can focus on clinically relevant contrast regions and reduce the impact of pixel misalignment, enhance the accuracy of blood vessel structures, and make the generated images have more accurate blood vessel structures.

[0018] In a possible implementation manner, the calculation formula of the loss function is: , , , , , , In the formula, represents the said loss function; , , and all represent weight factors; represents element-wise multiplication; represents the target PatchNCE loss function; represents the source PatchNCE loss function; represents the attention loss function; represents the combined discriminator loss function; represents the number of data in the said dataset; represents the anchor point supplement determined on the image output by the said conversion device during parameter optimization; Represents the positive sample patch defined on the input image of the conversion device during parameter optimization; Represents the negative sample patch defined on the input image of the conversion device during parameter optimization; Represents the th attention map generated by the contrast enhancement attention guidance sub-module during parameter optimization; Represents the denoising result obtained by the Bayesian posterior sampling sub-module during parameter optimization; Represents the binary contrast enhancement mask; Represents the InfoNCE loss function; Represents the non-saturated adversarial loss function with gradient penalty.

[0019] The contrast-aware technology corresponding to the above loss function, after being combined with the binary contrast enhancement mask, can enhance the accuracy of the vascular structure on the basis of focusing on clinically relevant contrast regions and reducing the influence of pixel misalignment, making the generated image have a more accurate vascular structure.

[0020] Another technical solution of the present invention is to provide a mask-guided contrast agent-free angiography generation method, including the following steps: S1: Use the parameter optimization device to generate a binary contrast enhancement mask by using a data set, and substitute the binary contrast enhancement mask into the loss function for highlighting the vascular region affected by the contrast agent to optimize the parameters of the conversion device through the loss function, obtain the model parameters and transmit them to the conversion device; S2: Use the noise generation device to generate a Gaussian noise image by a random Gaussian noise iteration algorithm, and obtain a contrast agent-free computed tomography image by the acquisition device; S3: Use the conversion device to integrate the Gaussian noise image into the contrast agent-free computed tomography image, and obtain a computed tomography angiography through a fusion method of feature sampling and step-by-step denoising.

[0021] The method for generating contrast-agent-free angiography based on mask guidance disclosed by the present invention first uses a parameter optimization device to optimize the parameters of the conversion device by using a loss function that can highlight the influence of the contrast agent, which improves the ability of the conversion device to capture complex image details. Subsequently, the Gaussian noise image is integrated into the contrast-agent-free computed tomography image through the conversion device, and the computed tomography angiography is obtained through the fusion method of feature sampling and step-by-step denoising. Finally, it is ensured that the obtained computed tomography angiography has sufficient consistency and realism. This process of integrating the loss function that can highlight the influence of the contrast agent into the diffusion process based on the Gaussian noise image generated by the random Gaussian noise iteration algorithm not only effectively promotes the conversion of the contrast-agent-free computed tomography image into a higher-quality computed tomography angiography, but also stabilizes the parameter optimization process, improves the sharpness and fidelity of the generated vascular structure, and thus overcomes the technical defects of the existing RegGAN model architecture.

[0022] In a possible implementation manner, the step S1 includes the following steps: S101: Retrieve multiple groups of contrast-agent-free computed tomography images and their corresponding computed tomography angiographies from the database to obtain the data set; S102: Input all the data in the data set into the parameter optimization device and the conversion device; S103: Use the unsupervised mask generation module in the parameter optimization device to generate a binary contrast enhancement mask by using the data set; S104: Process the data set through the encoder and decoder of each adversarial diffusion generation module in the conversion device to obtain the estimated denoised image corresponding to each adversarial diffusion generation module; S105: Use the contrast enhancement attention guidance sub-module in each adversarial diffusion generation module in the conversion device to perform element-wise multiplication and ReLU activation on the estimated denoised image and the binary contrast enhancement mask in sequence to generate the attention map corresponding to each adversarial diffusion generation module; S106: Respectively use the attention map obtained in step S105 by each contrast enhancement attention guidance sub-module to optimize the parameters of the encoder and decoder in the adversarial diffusion generation module where it is located; S107: Respectively use the Bayesian posterior sampling sub-module in each adversarial diffusion generation module in the conversion device to obtain the denoising result corresponding to each adversarial diffusion generation module according to the estimated denoised image; S108: Respectively use the angiography perception discrimination sub-module in each adversarial diffusion generation module in the conversion device to optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the adversarial diffusion generation module where it is located according to the denoising result obtained in step S107; S109: Substitute the binary contrast enhancement mask into the loss function through the mask-guided feature alignment module in the parameter optimization device, so as to optimize the parameters of the conversion device using the loss function, obtain model parameters, and transmit them to all the encoders, the decoder, and the Bayesian posterior sampling sub-module.

[0023] In the above parameter optimization process, by introducing and integrating the binary contrast enhancement mask, not only the original loss function is extended, highlighting the regions in the finally obtained image of the conversion device that are most significantly affected by the contrast agent, but also it can guide the conversion device to prioritize the blood vessel regions most affected by the contrast agent, thereby improving the accuracy and continuity of these key anatomical features. Description of the Drawings

[0024] Figure 1 Schematic structural diagram of a mask-guided contrast-agent-free angiography generation system disclosed in an embodiment of the present invention; Figure 2 Operating principle diagram of the conversion device disclosed in an embodiment of the present invention; Figure 3 Schematic structural diagram of the adversarial diffusion generation module disclosed in an embodiment of the present invention; Figure 4 Principle diagram of the parameter optimization device disclosed in an embodiment of the present invention; Figure 5 Method flow chart disclosed in an embodiment of the present invention; Figure 6 Flow chart of step S1 disclosed in an embodiment of the present invention. Detailed Embodiments

[0025] First of all, those skilled in the art should understand that these embodiments are only used to explain the technical principles of the embodiments of this application, and are not intended to limit the protection scope of the embodiments of this application. Those skilled in the art can make adjustments according to needs to adapt to specific application scenarios.

[0026] In the description of the embodiments of this application, it should be noted that unless otherwise clearly specified and limited, the terms "electrically connected" and "establish an electrical connection relationship" should be understood in a broad sense, that is, it should be understood that both or more of them have an electrical relationship, which can be achieved through wires, or can be a radio connection, or a combination of the two; it can be directly connected, or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of this application can be understood according to specific situations.

[0027] In the description of the embodiments of the present application, it should be noted that unless otherwise clearly specified and limited, whenever it is mentioned that a certain module or device is parameter-optimized by or using a certain loss function, it generally refers to a series of processes in which the model parameters of this module or device are taken as the parameters to be solved, and the goal is to obtain the optimal value of the loss function. Through optimization model solving algorithms (such as particle swarm optimization algorithm, greedy algorithm, Newton's method, back-gradient method), the specific values of the parameters to be solved are obtained, and the specific values of the parameters to be solved are used as the model parameters of this module or device.

[0028] In the embodiments of the present application, unless otherwise clearly specified and limited, the first feature being "on" or "under" the second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature being "above", "over" and "on top of" the second feature may be that the first feature is directly above or obliquely above the second feature, or merely indicates that the first feature has a higher horizontal height than the second feature. The first feature being "under", "beneath" and "underneath" the second feature may be that the first feature is directly below or obliquely below the second feature, or merely indicates that the first feature has a lower horizontal height than the second feature.

[0029] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] See Figures 1 to 6 , the embodiments of the present application disclose a mask-guided contrast-agent-free angiography generation system. Figure 1 FIG. is a schematic structural diagram of the angiography generation system. The angiography generation system includes a noise generation device, a collection device, a conversion device, and a parameter optimization device. Among them, the conversion device is electrically connected to the noise generation device and the collection device at the same time, and the parameter optimization device is electrically connected to the conversion device. The collection device is configured to obtain contrast-agent-free computed tomography images (in all figures, the contrast-agent-free computed tomography images are simply referred to as NCCT images). The single arrows in the figure represent that data can only be propagated unidirectionally in the direction of the arrow, and the double arrows represent that data can be propagated bidirectionally between the two connected devices.

[0031] In this angiography generation system, the noise generation device is configured to generate a Gaussian noise image with a random Gaussian noise iteration algorithm, that is, pure Gaussian noise. Specifically, in this embodiment, the iteration formula of the random Gaussian noise iteration algorithm is as follows: , , , where, is a composite number, representing the initially determined diffusion steps; is a composite number The root cause factor, i.e., except 1 and represents the iterative diffusion step size; represents the noise at the th iteration; represents the noise addition term at the th iteration; represents an arbitrarily small number; represents the conditional probability distribution of obtaining noise under the condition of obtaining noise ; represents the identity matrix; represents the Gaussian distribution; represents the minimum value of the exponential progress; represents the maximum value of the exponential progress.

[0032] In this angiography generation system, the conversion device is configured to integrate the Gaussian noise image into the non-contrast computed tomography image and obtain computed tomography angiography (abbreviated as CTA image in all figures) through the fusion of feature sampling and step-by-step denoising. Figure 2 For the operation schematic diagram of the conversion device, see Figure 2 , in this embodiment, the conversion device is a communication network structure composed of multiple cascaded adversarial diffusion generation modules, and the adversarial diffusion generation module is configured to convert the input image into the output image through the feature sampling method of encoding and decoding and the method of Bayesian rule sampling denoising. The number of adversarial diffusion generation modules should be the same as the number of steps of adding noise by the noise generation device. In this embodiment, 4 adversarial diffusion generation modules are adopted, and each adversarial diffusion generation module corresponds to a denoising step of a diffusion model.

[0033] See Figure 3 and Figure 4, in this embodiment, the anti-diffusion generation module includes an encoder, a decoder, a Bayesian posterior sampling sub-module, a contrast-enhanced attention guidance sub-module, and a contrast perception discrimination sub-module; the input end of the encoder in the anti-diffusion generation module at the first position is electrically connected to the output ends of both the noise generation device and the acquisition device, and the input end of the encoder in the remaining anti-diffusion generation modules is electrically connected to the output end of the Bayesian posterior sampling sub-module in the previous anti-diffusion generation module. In each anti-diffusion generation module, the input end of the decoder is electrically connected to the output end of the encoder, the input end of the Bayesian posterior sampling sub-module is electrically connected to the output end of the decoder, the input end of the contrast perception discrimination sub-module is electrically connected to the output ends of both the Bayesian posterior sampling sub-module and the parameter optimization device, the input end of the contrast-enhanced attention guidance sub-module is electrically connected to the output ends of both the decoder and the parameter optimization device, and the input end of the encoder is also electrically connected to the output ends of both the contrast perception discrimination sub-module and the contrast-enhanced attention guidance sub-module.

[0034] In the anti-diffusion generation module, the encoder is configured to perform a convolution operation on its input image (if it is the encoder in the anti-diffusion generation module at the first position, the input image is a Gaussian noise image and a non-contrast-enhanced computed tomography image) to extract anatomical structure and texture information, and incorporate temporal information during the extraction process, and output an encoded image; the decoder is configured to perform deconvolution denoising on the encoded image to obtain an estimated denoised image; the Bayesian posterior sampling sub-module is configured to denoise the estimated denoised image through Bayesian sampling rules to obtain a denoising result; the contrast-enhanced attention guidance sub-module is configured to, after the parameter optimization device is started, multiply the estimated denoised image and the binary contrast-enhanced mask element by element and perform ReLU activation in sequence to generate an attention map, and then optimize the parameters of the encoder and decoder in the anti-diffusion generation module where it is located with the goal of obtaining the minimum value of the distance between the obtained attention map and the binary contrast-enhanced mask; the contrast perception discrimination sub-module is configured to, after the parameter optimization device is started, optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the anti-diffusion generation module where it is located with the goal of obtaining the minimum value of the non-saturating adversarial loss function with gradient penalty.

[0035] In this angiography generation system, the parameter optimization device is configured to generate a binary contrast-enhanced mask based on a subtraction strategy using a data set, substitute the binary contrast-enhanced mask into a loss function for highlighting the blood vessel region affected by the contrast agent, and optimize the parameters of the conversion device using the loss function to obtain model parameters and transmit them to the conversion device.

[0036] See Figure 4, in this embodiment, the parameter optimization device includes an unsupervised mask generation module and a mask-guided feature alignment module. Among them, the output end of the unsupervised mask generation module is electrically connected to the input ends of all contrast enhancement attention guidance sub-modules and the input end of the angiography perception discrimination sub-module at the same time. The input end of the mask-guided feature alignment module is electrically connected to the output end of the unsupervised mask generation module and the output end of the Bayesian posterior sampling sub-module in the adversarial diffusion generation module located at the end at the same time.

[0037] In the parameter optimization device, the unsupervised mask generation module is configured to generate a binary contrast enhancement mask based on a subtraction strategy using a data set. In this embodiment, the calculation formula of the subtraction strategy is: , where, represents the binary contrast enhancement mask of the th data in the data set; represents the computed tomography angiography of the th data in the data set; represents the non-contrast computed tomography image of the th data in the data set; represents binarization processing. In this embodiment, its definition formula is: , wherein represents the threshold set by the threshold technology verified by empirical experiments, and takes the value of 20 in this embodiment.

[0038] In the parameter optimization device, the mask-guided feature alignment module is configured to substitute the binary contrast enhancement mask into the loss function to optimize the parameters of the conversion device using the loss function, obtain the model parameters and transmit them to all the encoders, decoders and Bayesian posterior sampling sub-modules. In this embodiment, the calculation formula of the loss function is: , , , , , , where, represents the loss function; , , and both represent weight factors; in this embodiment, , , , ; represents element-wise multiplication; represents the target PatchNCE loss function; represents the source PatchNCE loss function; represents the attention loss function; represents the combined discriminator loss function; represents the number of data in the dataset; represents the anchor patch determined on the image output by the conversion device during parameter optimization; represents the positive sample patch defined on the input image of the conversion device during parameter optimization; represents the negative sample patch defined on the input image of the conversion device during parameter optimization; represents the th attention map generated by the contrast enhancement attention guidance sub-module during parameter optimization; represents the denoising result obtained by the Bayesian posterior sampling sub-module during parameter optimization; represents the binary contrast enhancement mask; represents the InfoNCE loss function; represents the non-saturated adversarial loss function with gradient penalty.

[0039] Next, the usage method of the mask-guided non-contrast angiography generation system in this embodiment will be further disclosed. The flowchart of this method is as shown in Figure 5 and the method includes the following steps: S1: Use the parameter optimization device to generate a binary contrast enhancement mask using the dataset, and substitute the binary contrast enhancement mask into the loss function for highlighting the blood vessel area affected by the contrast agent to optimize the parameters of the conversion device through the loss function, obtain the model parameters, and transmit them to the conversion device.

[0040] See Figure 4 and Figure 6, in this embodiment, step S1 includes the following steps: S101: Retrieve multiple groups of non-contrast computed tomography images and their corresponding computed tomography angiographies from the database to obtain a dataset.

[0041] S102: Input all the data in the dataset into the parameter optimization device and the conversion device.

[0042] S103: Use the unsupervised mask generation module in the parameter optimization device to generate a binary contrast enhancement mask using the dataset.

[0043] S104: Process the dataset through the encoders and decoders of each adversarial diffusion generation module in the conversion device to obtain the estimated denoised images corresponding to each adversarial diffusion generation module. In this process, the estimated denoised image corresponding to an adversarial diffusion generation module is a function of the model parameters of the encoder and decoder of that adversarial diffusion generation module.

[0044] S105: Use the contrast enhancement attention guidance sub-module of each adversarial diffusion generation module in the conversion device to multiply the estimated denoised image and the binary contrast enhancement mask element by element and then perform ReLU activation in sequence to generate the attention maps corresponding to each adversarial diffusion generation module. The attention map corresponding to an adversarial diffusion generation module is a function of the model parameters of the encoder and decoder of that adversarial diffusion generation module.

[0045] S106: Respectively, use the attention maps obtained in step S105 by each contrast enhancement attention guidance sub-module to optimize the parameters of the encoder and decoder in the adversarial diffusion generation module where it is located. The specific method is: optimize the parameters of the encoder and decoder in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum distance between the attention map and the binary contrast enhancement mask.

[0046] S107: Respectively, use the Bayesian posterior sampling sub-module of each adversarial diffusion generation module in the conversion device to obtain the denoising results corresponding to each adversarial diffusion generation module based on the estimated denoised image. The denoising result corresponding to an adversarial diffusion generation module is a function of the model parameters of the encoder, decoder, and Bayesian posterior sampling sub-module of that adversarial diffusion generation module's own auction house.

[0047] S108: Respectively, use the angiography perception discrimination sub-module of each adversarial diffusion generation module in the conversion device to optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the adversarial diffusion generation module where it is located based on the denoising results obtained in step S107. The specific method is: substitute the denoising result into the non-saturating adversarial loss function with gradient penalty, and optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum value of the non-saturating adversarial loss function with gradient penalty.

[0048] S109: Substitute the binary contrast enhancement mask into the loss function through the mask-guided feature alignment module in the parameter optimization device to optimize the parameters of the conversion device using the loss function, obtain the model parameters, and transmit them to all the encoders, decoders, and Bayesian posterior sampling sub-modules.

[0049] During the parameter optimization process, the coefficient is determined as: , which balances gradient stability and detail retention according to the ablation study. In this embodiment, data preprocessing is also performed when obtaining the dataset in step S101. For data preprocessing, the ratio of window width to window level is set according to clinical guidelines; computed tomography angiography is adjusted to a vascular window [600, 200] to enhance contrast visualization, while non-contrast computed tomography images use a mediastinal window [400, 50] to distinguish soft tissues, and at the same time, rigid affine registration based on mutual information maximization is implemented using SimpleITK.

[0050] S2: Generate a Gaussian noise image through the noise generation device using a random Gaussian noise iteration algorithm, and obtain a non-contrast computed tomography image through the acquisition device.

[0051] S3: Incorporate the Gaussian noise image into the non-contrast computed tomography image through the conversion device, and obtain computed tomography angiography through a fusion method of feature sampling and progressive denoising.

[0052] In this embodiment, the iterative diffusion step size , the noise schedule is defined by , and the total number of diffusion steps . For the mask-guided PatchNCE loss function, patch embedding uses cosine similarity comparison, and the temperature . The conversion device was trained for 50 epochs (about 120 hours) on four NVIDIA RTX 4090 GPUs. The dataset split strictly follows patient-level separation: 100 patients are used for training, 25 patients are used for validation, and 25 patients are used for testing to ensure no data leakage. Inference maintains the same diffusion parameters as training to retain the random dynamics.

[0053] Next, the technical effects of the mask-guided non-contrast angiography generation system in this embodiment will be described in detail. In this embodiment, this angiography generation system is compared with existing image-to-image conversion devices. The image-to-image conversion devices used include GAN-based frameworks such as Pix2Pix, CycleGAN, CTA-GAN, UNIT, MUNIT, and NICE-GAN, as well as the diffusion-based model SynDiff.

[0054] The quantitative evaluation results show that the angiography generation system is superior to these devices in multiple metrics. Specifically, the PSNR value of the angiography generation system is 24.74, showing a significant improvement compared to Pix2Pix (21.28), CycleGAN (22.97), and other methods such as CTA-GAN (23.25) and UNIT (22.60). This indicates that the angiography generation system performs excellently in restoring image details and fidelity, and the generated images are of higher quality and closer to the target images in terms of quality. In terms of the SSIM metric, the angiography generation system reaches 86.43. Compared to Pix2Pix (74.44) and CycleGAN (80.42), the angiography generation system exhibits better structural similarity, indicating that the generated images are not only highly similar to the target images in terms of overall structure but also perform better in texture and detail retention. In terms of the NMAE metric, the value of the angiography generation system is 0.0317, lower than that of Pix2Pix (0.0453), CycleGAN (0.0382), and other comparison methods, indicating that it is superior in terms of brightness and intensity alignment with the target images. This further demonstrates that the angiography generation system has powerful performance in maintaining brightness consistency and detail reconstruction. In terms of the MSSSIM metric, the angiography generation system reaches 92.66, significantly higher than CycleGAN (76.52) and CTA-GAN (79.71). This demonstrates the ability of the angiography generation system to retain multi-scale structural features and provide higher image quality at multiple resolution levels. The improvement in MSSSIM highlights the ability of the angiography generation system to effectively capture and maintain structural details, which is crucial for accurate visual interpretation. In terms of the MI metric, the angiography generation system reaches 1.325, surpassing SynDiff (1.243) and CycleGAN (1.237). This indicates that the angiography generation system performs better in retaining the mutual information between the generated images and the target images, reflecting its superior ability to retain important image features and information during the generation process.

[0055] These quantitative results clearly show that the angiography generation system has good beneficial effects, especially in terms of PSNR, SSIM, NMAE, MSSSIM, and MI, fully verifying its excellent performance in detail retention and image consistency.

[0056] In summary, the contrast-agent-free angiography generation system based on mask guidance disclosed in this embodiment can, by setting a noise generation device, an acquisition device, a conversion device, and a parameter optimization device, use the loss function that can highlight the influence of the contrast agent by the parameter optimization device to optimize the parameters of the conversion device, thereby enhancing the ability of the conversion device to capture complex image details. Subsequently, the Gaussian noise image can be integrated into the contrast-agent-free computed tomography image by the conversion device, and the computed tomography angiography can be obtained through the fusion method of feature sampling and gradual denoising, ensuring that the obtained computed tomography angiography has sufficient consistency and realism. This integration of the loss function that can highlight the influence of the contrast agent into the diffusion process based on the Gaussian noise image generated by the random Gaussian noise iteration algorithm not only effectively ensures the conversion of the contrast-agent-free computed tomography image into a high-quality computed tomography angiography, but also stabilizes the parameter optimization process, improves the sharpness and fidelity of the generated vascular structure, and thus overcomes the technical defects of the existing RegGAN model architecture.

[0057] In the description of the embodiments of the present application, it should be noted that in the description of the present application, the terms "inner", "outer", etc., indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present application.

[0058] In the description of the present application, the descriptions with reference to terms such as "one embodiment", "some embodiments", "in this embodiment", "specific examples", or "some examples" mean that the specific features, mechanisms, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0059] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A mask-guided contrast agent-free angiography generation system, characterized in that Comprising: A noise generation device, configured to generate a Gaussian noise image by a random Gaussian noise iteration algorithm; An acquisition device, configured to acquire a non-contrast computed tomography image; A conversion device, electrically connected to the noise generation device and the acquisition device at the same time, configured to integrate the Gaussian noise image into the non-contrast computed tomography image, and obtain computed tomography angiography through a fusion method of feature sampling and step-by-step denoising; A parameter optimization device, electrically connected to the conversion device, configured to generate a binary contrast enhancement mask based on a subtraction strategy using a data set, substitute the binary contrast enhancement mask into a loss function for highlighting the vascular region affected by the contrast agent, and optimize the parameters of the conversion device using the loss function, obtain model parameters and transmit them to the conversion device.

2. The mask-guided contrast-agent-free angiography generation system according to claim 1, wherein The iteration formula of the random Gaussian noise iteration algorithm is as follows: ， ， ， In the formula, is a composite number, representing the initially proposed number of diffusion steps; is a composite number prime factor thereof; Represent the noise at the step of the iteration; Denote the noise addition term for the iteration step represents any arbitrarily small number; Represents the conditional probability distribution of obtaining noise under the condition of obtaining noise ; represents the identity matrix; represents a Gaussian distribution; Represents the minimum value of the exponential progress; Represents the maximum value of the exponential progress.

3. The mask-guided contrast-agent-free angiography generation system according to claim 1 or 2, characterized in that, The conversion device is a communication network structure formed by connecting multiple adversarial diffusion generation modules in series, and the adversarial diffusion generation module is configured to convert an input image into an output image through a feature sampling method of encoding and decoding and a denoising method of Bayesian rule sampling.

4. The mask-guided contrast-agent-free angiography generation system according to claim 3, wherein The adversarial diffusion generation module includes an encoder, a decoder, a Bayesian posterior sampling sub-module, a contrast enhancement attention guidance sub-module, and a contrast agent perception discriminant sub-module; The input end of the encoder in the adversarial diffusion generation module at the first position is electrically connected to the output end of the noise generation device and the output end of the acquisition device at the same time, and the input end of the encoder in the remaining adversarial diffusion generation modules is electrically connected to the output end of the Bayesian posterior sampling sub-module in the previous adversarial diffusion generation module; In each adversarial diffusion generation module, the input end of the decoder is electrically connected to the output end of the encoder, the input end of the Bayesian posterior sampling sub-module is electrically connected to the output end of the decoder, the input end of the contrast agent perception discriminant sub-module is electrically connected to the output end of the Bayesian posterior sampling sub-module and the output end of the parameter optimization device at the same time, the input end of the contrast enhancement attention guidance sub-module is electrically connected to the output end of the decoder and the output end of the parameter optimization device at the same time, and the input end of the encoder is also electrically connected to the output end of the contrast agent perception discriminant sub-module and the output end of the contrast enhancement attention guidance sub-module at the same time.

5. The mask-guided contrast-agent-free angiography generation system according to claim 4, wherein The encoder is configured to perform a convolution operation on its input image to extract anatomical structure and texture information, and integrate temporal information during the extraction process, and output an encoded image; The decoder is configured to perform deconvolution denoising on the encoded image to obtain an estimated denoised image; The Bayesian posterior sampling sub-module is configured to denoise the estimated denoised image through a Bayesian sampling rule to obtain a denoising result; The contrast-enhanced attention guidance sub-module is configured to, after the parameter optimization device is started, perform element-wise multiplication and ReLU activation on the estimated denoised image and the binary contrast-enhanced mask in sequence to generate an attention map, and optimize the parameters of the encoder and decoder in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum value of the distance between the attention map and the binary contrast-enhanced mask; The angiography-aware discrimination sub-module is configured to, after the parameter optimization device is started, optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the adversarial diffusion generation module where it is located with the goal of obtaining the minimum value of the non-saturated adversarial loss function with gradient penalty; 6. The mask-guided contrast-agent-free angiography generation system according to claim 5, wherein The parameter optimization device includes: An unsupervised mask generation module, whose output is electrically connected to the input ends of all the contrast-enhanced attention guidance sub-modules and the input end of the angiography-aware discrimination sub-module at the same time, and is configured to generate a binary contrast-enhanced mask based on a subtraction strategy by using the data set; A mask-guided feature alignment module, whose input end is electrically connected to the output end of the unsupervised mask generation module and the output end of the Bayesian posterior sampling sub-module in the adversarial diffusion generation module at the end at the same time, and is configured to substitute the binary contrast-enhanced mask into the loss function to optimize the parameters of the conversion device by using the loss function, obtain model parameters, and transmit them to all the encoders, the decoders, and the Bayesian posterior sampling sub-modules; 7. The mask-guided contrast-agent-free angiography generation system according to claim 6, wherein The calculation formula of the subtraction strategy is: ， wherein, represent the binary contrast enhancement mask for the th data in the dataset; representing computed tomography angiography of the n-th data in the dataset; Representing the non-contrast computed tomography image of the Represents binarization processing.

8. The mask-guided contrast-agent-free angiography generation system according to claim 7, wherein The calculation formula of the loss function is: ， ， ， ， ， ， wherein, representing the loss function; , , and all represent weighting factors; represents element-wise multiplication; Represents the target PatchNCE loss function; Represents the source PatchNCE loss function; Represents an attention loss function; Represents the combined discriminator loss function; represents the number of data in the dataset; Denotes the anchor point supplement determined on the image output by the conversion device during parameter optimization; Represents positive sample patches defined on the input image of the conversion device during parameter optimization; Represents negative sample patches defined on the input image of the conversion device when performing parameter optimization; The attention map generated by the contrast enhancement attention guidance sub-module when performing parameter optimization; It represents the denoising result obtained by the Bayesian posterior sampling sub-module during parameter optimization; representing the binary contrast enhancement mask; Represents the InfoNCE loss function; Represents a non-saturated adversarial loss function with gradient penalty.

9. A method for generating mask-guided contrast agent-free angiography, characterized in that, Applied to the mask-guided contrast agent-free angiography generation system according to any one of claims 1-8, the method includes the following steps: S1: Use the parameter optimization device to generate a binary contrast-enhanced mask by using a data set, and substitute the binary contrast-enhanced mask into a loss function for highlighting the vascular region affected by the contrast agent to optimize the parameters of the conversion device by using the loss function, obtain model parameters, and transmit them to the conversion device; S2: Use the noise generation device to generate a Gaussian noise image by using a random Gaussian noise iteration algorithm, and acquire a contrast agent-free computed tomography image by using an acquisition device; S3: Use the conversion device to integrate the Gaussian noise image into the contrast agent-free computed tomography image, and obtain a computed tomography angiography through a fusion method of feature sampling and progressive denoising; 10. The method for generating mask-guided contrast agent-free angiography according to claim 9, wherein, The step S1 includes the following steps: S101: Retrieve multiple groups of contrast agent-free computed tomography images and their corresponding computed tomography angiographies in a database to obtain the data set; S102: Input all the data in the data set into the parameter optimization device and the conversion device; S103: Use the unsupervised mask generation module in the parameter optimization device to generate a binary contrast-enhanced mask by using the data set; S104: Obtain the estimated denoised image corresponding to each adversarial diffusion generation module through the processing of the data set by the encoder and decoder of each adversarial diffusion generation module in the conversion device; S105: Multiply the estimated denoised image and the binary contrast enhancement mask element by element and perform ReLU activation through the contrast enhancement attention guidance sub-module of each pair of counter-diffusion generation modules in the conversion device to generate an attention map corresponding to each counter-diffusion generation module; S106: Respectively, use the attention map obtained in step S105 by each contrast enhancement attention guidance sub-module to optimize the parameters of the encoder and decoder in the counter-diffusion generation module where it is located; S107: Respectively, obtain the denoising result corresponding to each counter-diffusion generation module according to the estimated denoised image through the Bayesian posterior sampling sub-module of each counter-diffusion generation module in the conversion device; S108: Respectively, optimize the parameters of the encoder, decoder, and Bayesian posterior sampling sub-module in the counter-diffusion generation module where it is located according to the denoising result obtained in step S107 through the angiography perception discrimination sub-module of each counter-diffusion generation module in the conversion device; S109: Substitute the binary contrast enhancement mask into the loss function through the mask-guided feature alignment module in the parameter optimization device to optimize the parameters of the conversion device using the loss function, obtain model parameters, and transmit them to all the encoders, decoders, and Bayesian posterior sampling sub-modules.

Images