Low-dose CTA image denoising model training method, denoising method and device

Abstract

The application provides a low-dose CTA image denoising model training method, a denoising method and equipment. The method comprises the following steps: obtaining cardiovascular and edge noise enhancement data of a low-dose CTA image based on a cardiovascular segmentation model and an edge extraction model; and generating a generative adversarial network based on a target loss training cycle according to the data and a conventional-dose CTA image, so as to train the generative adversarial network into a low-dose CTA image denoising model. The target loss comprises a cardiovascular preservation loss obtained based on the cardiovascular segmentation model and an edge preservation loss obtained based on the edge extraction model. The application can combine the enhancement of blood vessels and edges with the global denoising of a low-dose CTA image in the training process of a low-dose CTA image denoising model, so that the model can take into account the preservation of subtle structures of an image and the improvement of overall image quality, thereby effectively improving the application effectiveness and reliability of a low-dose CTA image denoising result obtained by using the model, and improving the reliability of CTA clinical diagnosis.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, and in particular to a training method, denoising method and device for low-dose CTA image denoising models. Background Technology

[0002] Coronary CT angiography (CT) is an important tool for diagnosing cardiovascular diseases, offering advantages such as non-invasiveness and high resolution. However, adhering to the "As Low As Reasonably Achievable" (ALARA) principle requires performing CTA (computed tomography angiography) examinations with low radiation doses and reduced contrast agent usage. This results in significant noise and blurring in the low-dose CTA images, especially in vascular regions. Noise not only reduces vascular clarity but can also lead to missed diagnoses of clinically significant microscopic lesions such as vascular stenosis and plaque occlusion. Therefore, designing noise reduction methods for low-dose CTA images is a widely discussed issue in this field.

[0003] Currently, deep learning-based medical image denoising technology has made significant progress in the field of image denoising, especially demonstrating great potential in handling complex structures and noise characteristics. Deep learning denoising methods for low-dose CTA images are mainly divided into supervised learning and self-supervised learning. Due to difficulties in obtaining paired data and the inability of simulated noise to represent the complexity of real noise during supervised learning, researchers have begun to turn to self-supervised learning. Self-supervised learning constructs self-supervised tasks to achieve denoising of low-dose CTA images without a noise-free reference image. For example, some researchers have introduced CycleGAN for low-dose CTA image denoising, using unpaired low-dose CTA images and regular-dose CTA images for network training. This maps the low-dose CTA images into the regular-dose CTA image domain, effectively reducing the requirements for data acquisition and eliminating the need for pre-assumptions about noise in the low-dose CTA images.

[0004] However, existing low-dose CTA image denoising methods based on self-supervised learning often suffer from "content inconsistency" issues because CycleGAN relies on unpaired image pairs for training. While deep learning techniques have made significant progress in low-dose CTA image denoising, most methods still focus on global denoising, neglecting specific enhancements of fine structures such as blood vessels. This can lead to the loss of crucial anatomical information during the denoising process.

[0005] Therefore, there is an urgent need to design a low-dose CTA image denoising method that can enhance the cardiovascular features of low-dose CTA images in a targeted manner while performing overall denoising on low-dose CTA images. Summary of the Invention

[0006] In view of this, embodiments of this application provide a low-dose CTA image denoising model training method, denoising method, and device to eliminate or improve one or more defects existing in the prior art.

[0007] One aspect of this application provides a method for training a low-dose CTA image denoising model, comprising:

[0008] Cardiovascular and edge noise enhancement data corresponding to low-dose CTA images were obtained based on cardiovascular segmentation and edge extraction models.

[0009] Based on the cardiovascular and edge noise enhancement data and the preset conventional dose CTA images, a preset recurrent generative adversarial network is trained based on the target loss to train the recurrent generative adversarial network into a low-dose CTA image denoising model. The target loss includes cardiovascular preservation loss obtained based on the cardiovascular segmentation model and edge preservation loss obtained based on the edge extraction model. The low-dose CTA image corresponds to a radiation dose less than that of the conventional dose CTA image.

[0010] In some embodiments of this application, the acquisition of cardiovascular and edge noise enhancement data corresponding to low-dose CTA images based on cardiovascular segmentation models and edge extraction models includes:

[0011] Multiple data pairs were acquired, each containing a low-dose CTA image and a standard-dose CTA image.

[0012] Based on the cardiovascular segmentation model, the cardiovascular binary mask corresponding to each of the low-dose CTA images is extracted respectively, and based on the edge extraction model, the edge binary mask corresponding to each of the low-dose CTA images is extracted respectively.

[0013] Based on the cardiovascular binary mask and edge binary mask corresponding to each of the low-dose CTA images, the cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images are obtained respectively;

[0014] Each low-dose CTA image is combined with its corresponding cardiovascular and edge noise enhancement data and the conventional-dose CTA image in the data pair containing the low-dose CTA image to form a data sample. Each data sample contains one cardiovascular and edge noise enhancement data and one conventional-dose CTA image.

[0015] In some embodiments of this application, obtaining the cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images based on the cardiovascular binary mask and edge binary mask corresponding to each low-dose CTA image includes:

[0016] The cardiovascular binary mask and edge binary mask corresponding to each low-dose CTA image are merged to obtain the key structure mask corresponding to each low-dose CTA image.

[0017] The regions containing the key structure masks in each of the low-dose CTA images are blurred, and random noise is added to each of the low-dose CTA images to obtain cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images.

[0018] In some embodiments of this application, the cyclic generative adversarial network includes: a first generative adversarial model and a second generative adversarial model;

[0019] The first generative adversarial model includes: a first discriminator, a first generator, and a second generator, wherein the output of the first generator is connected to the inputs of the second generator, the first discriminator, the cardiovascular segmentation model, and the edge extraction model, respectively.

[0020] The second generative adversarial model includes: a second discriminator, a second generator, and a first generator, wherein the output of the second generator is connected to the inputs of the first generator, the second discriminator, the cardiovascular segmentation model, and the edge extraction model, respectively.

[0021] Correspondingly, based on the cardiovascular and edge noise enhancement data and the preset conventional dose CTA images, a preset recurrent generative adversarial network is trained using a target loss to train the recurrent generative adversarial network into a low-dose CTA image denoising model. The target loss includes a cardiovascular preservation loss obtained based on the cardiovascular segmentation model and an edge preservation loss obtained based on the edge extraction model, including:

[0022] The data samples are used to perform at least one round of iterative training on the first generative adversarial model and the second generative adversarial model. In each iteration round, based on the target loss corresponding to the first generative adversarial model, a preset first training step is performed on the first generative adversarial model, and based on the target loss corresponding to the second generative adversarial model, a preset second training step is performed on the second generative adversarial model.

[0023] The first generative adversarial model, after iterative training, was determined as a low-dose CTA image denoising model.

[0024] In some embodiments of this application, the cardiovascular preservation loss includes: a first cardiovascular preservation loss corresponding to the first generative adversarial model; the edge preservation loss includes: a first edge preservation loss corresponding to the first generative adversarial model;

[0025] The target loss corresponding to the first generative adversarial model also includes: first identity preservation loss, first adversarial loss, and first cycle consistency loss;

[0026] Correspondingly, the step of performing a preset first training step on the first generative adversarial model based on the target loss corresponding to the first generative adversarial model includes:

[0027] The cardiovascular and edge noise enhancement data are input into a first generator so that the first generator outputs a first generated image corresponding to the low-dose CTA image, and a first identity preservation loss is obtained between the first generated image and the cardiovascular and edge noise enhancement data.

[0028] The first generated image is input into a first discriminator, a second generator, the cardiovascular segmentation model, and the edge extraction model, respectively, so that the first discriminator outputs a first adversarial loss between the first generated image and a preset conventional dose CTA image, the second generator outputs the noise reduction result data corresponding to the low dose CTA image, the cardiovascular segmentation model outputs a first cardiovascular preservation loss between the first generated image and the cardiovascular binary mask, and the edge extraction model outputs a first edge preservation loss between the first generated image and the edge binary mask;

[0029] The first cyclic consistency loss between the cardiovascular and edge noise enhancement data and the noise reduction result data is obtained;

[0030] The first generator is optimized based on the first identity preservation loss, the first adversarial loss, the first cardiovascular preservation loss, the first edge preservation loss, and the first cycle consistency loss; and the second generator is optimized based on the first cycle consistency loss.

[0031] In some embodiments of this application, the cardiovascular preservation loss further includes: a second cardiovascular preservation loss corresponding to the second generative adversarial model; the edge preservation loss further includes: a second edge preservation loss corresponding to the second generative adversarial model;

[0032] The target loss corresponding to the second generative adversarial model also includes: second identity preservation loss, second adversarial loss, and second cycle consistency loss;

[0033] Correspondingly, the second training step, based on the target loss corresponding to the second generative adversarial model, includes:

[0034] The conventional dose CTA image is input into a second generator so that the second generator outputs a second generated image corresponding to the conventional dose CTA image, and a second identity preservation loss between the second generated image and the conventional dose CTA image is obtained;

[0035] The second generated image is input into the second discriminator and the second generator respectively, so that the second discriminator outputs the second adversarial loss between the second generated image and the low-dose CTA image, and the second generator outputs the reconstruction result data corresponding to the conventional dose CTA image;

[0036] The second generated image and the conventional dose CTA image are input into the cardiovascular segmentation model so that the cardiovascular segmentation model outputs the corresponding second cardiovascular preservation loss, and the second generated image and the conventional dose CTA image are input into the edge extraction model so that the edge extraction model outputs the corresponding second edge preservation loss.

[0037] Acquire the second cyclic consistency loss between the conventional dose CTA image and the reconstructed result data;

[0038] The second generator is optimized based on the second identity preservation loss, the second adversarial loss, the second cardiovascular preservation loss, the second edge preservation loss, and the second cycle consistency loss; and the first generator is optimized based on the second cycle consistency loss.

[0039] In some embodiments of this application, both the first generator and the second generator are U-Net generators equipped with a high-frequency compensation module;

[0040] The U-Net generator is generated in advance by replacing all hop connections in the U-Net network with high-frequency compensation modules.

[0041] Another aspect of this application provides a low-dose CTA image denoising method, comprising:

[0042] The target low-dose CTA image is input into a preset low-dose CTA image denoising model, so that the low-dose CTA image denoising model outputs the denoising result data corresponding to the target low-dose CTA image. The low-dose CTA image denoising model is pre-trained based on the low-dose CTA image denoising model training method.

[0043] Output the noise reduction result data.

[0044] A third aspect of this application provides a training apparatus for a low-dose CTA image denoising model, comprising:

[0045] The data augmentation module is used to acquire cardiovascular and edge noise enhancement data corresponding to low-dose CTA images based on cardiovascular segmentation and edge extraction models;

[0046] The model training module is used to train a preset recurrent generative adversarial network based on the cardiovascular and edge noise enhancement data and the preset conventional dose CTA image, so as to train the recurrent generative adversarial network into a low-dose CTA image denoising model. The target loss includes cardiovascular preservation loss obtained based on the cardiovascular segmentation model and edge preservation loss obtained based on the edge extraction model. The low-dose CTA image corresponds to a radiation dose less than that of the conventional dose CTA image.

[0047] A fourth aspect of this application provides a low-dose CTA image noise reduction device, comprising:

[0048] The image denoising module is used to input the target low-dose CTA image into a preset low-dose CTA image denoising model, so that the low-dose CTA image denoising model outputs the denoising result data corresponding to the target low-dose CTA image, wherein the low-dose CTA image denoising model is pre-trained based on the low-dose CTA image denoising model training method.

[0049] The data output module is used to output the noise reduction result data.

[0050] A fifth aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the low-dose CTA image denoising model training method, and / or to implement the low-dose CTA image denoising method.

[0051] A sixth aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the low-dose CTA image denoising model training method, and / or implements the low-dose CTA image denoising method.

[0052] The seventh aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the low-dose CTA image denoising model training method, and / or implements the low-dose CTA image denoising method.

[0053] The low-dose CTA image denoising model training method provided in this application obtains cardiovascular and edge noise enhancement data corresponding to low-dose CTA images based on a cardiovascular segmentation model and an edge extraction model. Based on the cardiovascular and edge noise enhancement data and preset conventional-dose CTA images, a preset recurrent generative adversarial network (RGAN) is trained using a target loss to train the RGAN into a low-dose CTA image denoising model. The target loss includes a cardiovascular preservation loss obtained based on the cardiovascular segmentation model and an edge preservation loss obtained based on the edge extraction model. This method combines vascular and edge enhancement with global denoising of low-dose CTA images during the low-dose CTA image denoising model training process, enabling the model to balance the preservation of fine image structures with the improvement of overall image quality. This effectively improves the application effectiveness and reliability of the low-dose CTA image denoising results obtained using this model, thereby enhancing the reliability of CTA clinical diagnosis.

[0054] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0055] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings:

[0057] Figure 1 This is a schematic diagram of the first process of a low-dose CTA image denoising model training method in one embodiment of this application.

[0058] Figure 2 This is a schematic diagram of the second process of the low-dose CTA image denoising model training method in one embodiment of this application.

[0059] Figure 3 This is a schematic diagram of the third process of training a low-dose CTA image denoising model in one embodiment of this application.

[0060] Figure 4 This is a schematic diagram of the architecture of the first training step performed by the first generative adversarial model, cardiovascular segmentation model, and edge extraction model in one embodiment of this application.

[0061] Figure 5 This is a schematic diagram of the architecture of the second training step performed by the second generative adversarial model, cardiovascular segmentation model, and edge extraction model in one embodiment of this application.

[0062] Figure 6 This is a schematic diagram illustrating the optimization of the target loss corresponding to the first generative adversarial model in one embodiment of this application.

[0063] Figure 7 This is a schematic diagram illustrating the optimization of the target loss corresponding to the second generative adversarial model in one embodiment of this application.

[0064] Figure 8 This is a schematic diagram of the structure of the first generator and the second generator in one embodiment of this application.

[0065] Figure 9 This is a flowchart illustrating a low-dose CTA image denoising method according to an embodiment of this application.

[0066] Figure 10 This is a schematic diagram of the structure of a low-dose CTA image denoising model training device in one embodiment of this application.

[0067] Figure 11 This is a schematic diagram of the structure of a low-dose CTA image noise reduction device in one embodiment of this application. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0069] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0070] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0071] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0072] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0073] In recent years, deep learning-based medical image denoising technology has made significant progress in the field of image denoising, especially demonstrating great potential in handling complex structures and noise characteristics. Deep learning denoising methods for low-dose CTA images are mainly divided into two categories: supervised learning and self-supervised learning. In supervised learning, neural networks are typically trained on paired and aligned image pairs, achieving transfer from low-dose CT images to conventional-dose CT images by minimizing the reconstruction loss between low-dose and conventional-dose CT images. However, obtaining paired data is very difficult because it requires patients to remain still for extended periods to obtain perfectly aligned image data. To address this, some researchers have generated low-dose CT images by adding numerically simulated noise to clean images (i.e., conventional-dose CT images). Although this method has achieved some success in denoising, its limitation lies in the fact that simulated noise usually comes from simple mathematical models, while real-world noise often has complex statistical properties, leading to a degraded model performance when handling real-world noise.

[0074] To overcome the limitations of data acquisition, researchers have turned to self-supervised learning, which constructs self-supervised tasks to denoise low-dose CTA images without the need for noise-free reference images. While self-supervised methods effectively alleviate the need for paired images, these methods are often based on assumptions about the characteristics of noise and are difficult to handle situations where the characteristics of noise are unknown in real-world applications.

[0075] GAN-based denoising methods employ a game-like dynamics between the generator and discriminator, creating competition between them to ultimately optimize the image generation quality of the generator and the discrimination ability of the discriminator. During inference, a well-trained generator can produce high-quality images from input low-dose CT (low-dose CTA) images. To further improve detail preservation, researchers have proposed a generative adversarial network (GAN) combining Wasserstein distance and perceptual similarity for CT image denoising. This method effectively reduces noise while preserving more detail. However, GAN training often suffers from gradient vanishing and mode collapse issues. To address these problems, researchers introduced the WGAN-GP (Wasserstein GAN with Gradient Penalty) method, which alleviates gradient vanishing and mode collapse by adding a gradient penalty to the loss function, enabling the generator to produce more stable and high-quality denoised images. However, WGAN-GP training is complex and computationally expensive. Furthermore, Spectral Normalization GAN stabilizes the training process and improves the quality and stability of generated images by spectrally normalizing the discriminator weights. However, this approach may limit the expressive power of the model and requires a trade-off between stability and generation quality.

[0076] Furthermore, since the calculation of perceptual similarity relies on strictly paired images, this is difficult to obtain in actual CT examinations. To address this issue, researchers have proposed the unsupervised CycleGAN method, which learns the mapping between low-dose and high-dose CT images, achieving effective denoising of low-dose CTA images without the need for paired images. Specifically, CycleGAN can be trained using unpaired low-dose CTA images and regular-dose CTA image pairs, mapping the low-dose images to the high-dose image domain, effectively reducing the requirements for data acquisition and eliminating the need for pre-assumptions about noise in low-dose CTA images. However, because CycleGAN relies on unpaired image pairs for training, it lacks a clear mechanism to ensure the consistency of the content structure of the input images, often leading to content discrepancies between the denoised images and the original images. Although deep learning techniques have made significant progress in the field of low-dose CTA image denoising, most methods still focus on global denoising, neglecting specific enhancements of fine structures such as blood vessels, which may result in the loss of crucial anatomical information during the denoising process. Therefore, while performing overall noise reduction on low-dose CTA images, how to specifically enhance vascular features to improve the accuracy and reliability of low-dose CTA images in clinical diagnosis has become a critical issue that urgently needs to be addressed.

[0077] To avoid content inconsistency issues during Cycle GAN denoising, many subsequent studies have introduced different content constraint mechanisms. Some researchers have proposed a structural difference loss based on structural similarity (SSIM) to measure the difference between the generated and original images, optimizing the generator by minimizing this loss, allowing the model to focus more on preserving structural details while improving image quality. Similarly, some researchers have also introduced structural constraints into unsupervised MR-to-CT image synthesis using Cycle GAN, ensuring structural consistency during modality transformation by constraining modality-independent structural features. However, the structural similarity calculations of the above methods are usually performed through local window traversal, and the window size needs to be adjusted according to different structural scales, which increases computational complexity and adjustment difficulty.

[0078] Traditionally, researchers have relied primarily on various filter techniques to enhance vascular structures. The core idea behind these methods is to design specific mathematical models to distinguish blood vessels from background noise, thereby improving vascular visibility while simultaneously suppressing noise. Matched filters enhance vascular structures by convolving the image with a filter kernel matched to the morphology and intensity distribution of the blood vessels, offering advantages such as computational simplicity and orientation sensitivity. Gabor wavelet filters combine spatial and frequency domain characteristics, exhibiting orientation and scale selectivity, and can effectively extract the texture and edge features of blood vessels; however, they suffer from high computational complexity and difficulty in parameter tuning.

[0079] In the field of vessel enhancement, early research largely focused on the application and optimization of filters. Widely used image filtering methods based on Hessian matrix analysis utilize image gradient information and second derivatives to distinguish between tubular and non-tubular structures, thereby enhancing the saliency of vessels. With further research, many improved methods have been proposed, such as filters based on the Weingarten matrix and vessel enhancement diffusion techniques. These methods further optimize the extraction of vessel features, contributing to subsequent vessel enhancement tasks. However, past research primarily used vessel enhancement for subsequent vessel segmentation, without applying it to overall image denoising. To address this, some researchers have proposed a GAN-based CT image denoising method that combines vessel-specific filter loss with a controllable filter to extract vessel features from CT images, guiding the generator to focus more on preserving vessel features while denoising. However, filter-based vessel extraction methods often lack universality and require precise design of filter parameters.

[0080] Hessian matrix-based filters analyze the second derivative information of images, utilizing the eigenvalues ​​of the Hessian matrix and vascular measurement functions to identify and enhance blood vessels. They offer advantages such as orientation insensitivity and multi-scale adaptability, but are highly sensitive to noise and computationally intensive. Furthermore, these filters primarily enhance tubular structures like blood vessels, often attenuating important anatomical information in non-vascular regions. Therefore, achieving effective vascular enhancement while reducing noise in low-dose CTA images is crucial for improving image quality, enhancing lesion visibility, and increasing diagnostic accuracy.

[0081] Based on this, in order to perform overall denoising on low-dose CTA images while specifically enhancing the cardiovascular features of the images, embodiments of this application provide a low-dose CTA image denoising model training method, a low-dose CTA image denoising method, a low-dose CTA image denoising model training device for executing the low-dose CTA image denoising model training method, a low-dose CTA image denoising device for executing the low-dose CTA image denoising method, a physical device, a computer-readable storage medium, and a computer program product. These methods combine vascular and edge enhancement with global denoising of low-dose CTA images during the low-dose CTA image denoising model training process. While maintaining the overall image contour, global denoising of low-dose CTA images is performed, while targeted enhancement of blood vessels is applied, thereby improving the reliability of CTA clinical diagnosis.

[0082] The following examples will provide a detailed description.

[0083] Based on this, embodiments of this application provide a method for training a low-dose CTA image denoising model that can be implemented by a low-dose CTA image denoising model training device, see [link to relevant documentation]. Figure 1 The low-dose CTA image denoising model training method specifically includes the following:

[0084] Step 100: Obtain cardiovascular and edge noise enhancement data corresponding to low-dose CTA images based on cardiovascular segmentation and edge extraction models.

[0085] In one or more embodiments of this application, the cardiovascular segmentation model can also be referred to as a cardiovascular segmentation network, which can be implemented using a Res-Unet (Residual U-Net). The edge extraction model can also be referred to as an edge extraction network, which can be implemented using another Res-Unet (Residual U-Net).

[0086] Step 200: Based on the cardiovascular and edge noise enhancement data and the preset conventional dose CTA image, train a preset recurrent generative adversarial network based on the target loss, so as to train the recurrent generative adversarial network into a low-dose CTA image denoising model. The target loss includes cardiovascular preservation loss obtained based on the cardiovascular segmentation model and edge preservation loss obtained based on the edge extraction model. The low-dose CTA image corresponds to a radiation dose less than that of the conventional dose CTA image.

[0087] It is understood that the low-dose CTA image refers to the CTA image obtained after performing a CTA (computed tomography angiography) examination under the premise of low radiation dose and reduced contrast agent use. Low-dose CTA image can also be called computed tomography perfusion image (CTP image), and low-dose CTA image can also be abbreviated as LDCTA.

[0088] The conventional dose CTA image is a CTA image with a higher radiation dose compared to the low dose CTA image, and can also be called a computed tomography angiography image. The conventional dose CTA image can also be simply referred to as NDCTA.

[0089] For ease of explanation, in one or more embodiments of this application, the image domain of the low-dose CTA image can be identified as A, the image domain of the conventional-dose CTA image can be identified as B, and correspondingly, the cardiovascular and edge noise enhancement data can be identified as Aug.A.

[0090] It is understandable that, since the training of the recurrent generative adversarial network requires the use of cardiovascular preservation loss obtained from the cardiovascular segmentation model and edge preservation loss obtained from the edge extraction model, that is, the cardiovascular segmentation model and the edge extraction model are involved in the training of the recurrent generative adversarial network. Therefore, this application names the complete model architecture containing the recurrent generative adversarial network, the cardiovascular segmentation model and the edge extraction model as Contour-Preserving Denoising CycleGAN (EPD-CGAN). It combines vascular enhancement and global image denoising, and is a Cycle GAN network for denoising with targeted vascular enhancement and contour preservation.

[0091] The low-dose CTA image denoising model refers to a model used to denoise an input low-dose CTA image and output the corresponding denoising result data. The denoising result data is the corresponding conventional-dose CTA image. In one or more embodiments of this application, the dose refers to the radiation dose. Furthermore, the denoising result data can be identified using Rec.A.

[0092] It is understood that the target loss includes cardiovascular preservation loss obtained based on the cardiovascular segmentation model and edge preservation loss obtained based on the edge extraction model. It may also include identity preservation loss, adversarial loss and cycle consistency loss of the cyclic generative adversarial network itself. The specific settings can be made according to the actual application needs.

[0093] As can be seen from the above description, the low-dose CTA image denoising model training method provided in this application can combine blood vessel and edge enhancement with global denoising of low-dose CTA images during the low-dose CTA image denoising model training process. This allows the model to balance the preservation of the fine structure of the image with the improvement of the overall image quality, thereby effectively improving the application effectiveness and reliability of the low-dose CTA image denoising results obtained by using this model, and thus improving the reliability of CTA clinical diagnosis.

[0094] To further improve the effectiveness and reliability of cardiovascular and edge noise enhancement, a low-dose CTA image denoising model training method is provided in this application embodiment, see [link to relevant documentation]. Figure 2 Step 100 in the low-dose CTA image denoising model training method specifically includes the following:

[0095] Step 110: Acquire multiple data pairs, each containing a low-dose CTA image and a conventional-dose CTA image.

[0096] Specifically, as an unsupervised approach, the EPD-CGAN embodiment of this application aims to learn vascular enhancement and image denoising of low-dose CTA images using two sets of unpaired low-dose CTA images and conventional-dose CTA images, i.e., CTA images from different clinical patients without deformable registration.

[0097] In one example, this application embodiment uses a self-collected clinical dataset. This dataset contains low-dose CTA image data from 27 patients and conventional-dose CTA image data from 70 patients, all from the patients' chest regions. Specifically, there are a total of 8131 conventional-dose CTA images and 17150 low-dose CTA images. All images were acquired at a resolution of 512×512 pixels, focusing on detailed chest features. This application embodiment fixedly selected low-dose CTA images from 8 cases and conventional-dose CTA images from 3 cases to construct a test set containing 1500 low-dose CTA images and 874 conventional-dose CTA images. Since the conventional-dose and low-dose CTA images came from different patients, the dataset is unpaired.

[0098] In addition to the self-collected clinical dataset, this application embodiment also utilizes real clinical data from the AAPM-Mayo Clinic Low-Dose CT Challenge. This database contains ten sets of data, totaling 5936 slices. Specifically, each set includes full-dose images and synthetic quarter-dose (25%) images, provided in pairs at a 512×512 pixel resolution. The data sources cover the patient's abdomen, chest, and head. Therefore, this application embodiment uses full-dose images as the target and inputs the corresponding low-dose images into the neural network for training. Regarding the details of data preparation, this application embodiment uses L333 and L506 as the test set, and the remaining eight sets as the training and validation sets.

[0099] Furthermore, the ImageCAS dataset is specifically designed for high-precision coronary artery segmentation tasks, containing coronary CT images from multiple patients, providing detailed coronary artery annotations and high-quality image data. Specifically, this embodiment trains a vessel segmentation network and an edge segmentation network on the ImageCAS dataset. The vessel segmentation network aims to accurately delineate coronary artery structures, while the edge segmentation network is trained using binary edge results generated by the Canny algorithm as labels to capture subtle edge information in the images. The introduction of these segmentation networks aims to assist EPD-CGAN in better preserving key vascular and edge details during the conversion from low-dose CTA images to conventional-dose CTA images by providing additional structural and edge information.

[0100] Step 120: Extract the cardiovascular binary mask corresponding to each of the low-dose CTA images based on the cardiovascular segmentation model, and extract the edge binary mask corresponding to each of the low-dose CTA images based on the edge extraction model.

[0101] In the field of denoising, traditional data augmentation methods mainly involve noise injection, which simulates real-world imaging conditions by adding different types and levels of noise to the image. This allows the model to recover the original clear image under various noise environments, improving the model's denoising generalization ability. To improve the generalization of EPD-CGAN in terms of structure preservation, this application proposes a local personalized data augmentation technique to enhance the model's ability to recover key structures. First, two pre-trained Res-Unet frameworks are used to accurately segment the vascular structures and edges in the image, obtaining a cardiovascular binary mask S. v and edge binary mask S e :

[0102] S v =f vs (A)

[0103] S e =f es (A)

[0104] Among them, f vs Represents a cardiovascular segmentation model; f es () indicates the edge extraction model.

[0105] Step 130: Obtain the cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images based on the cardiovascular binary mask and edge binary mask corresponding to each of the low-dose CTA images.

[0106] Step 140: Combine the cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images with the conventional-dose CTA images in the data pair to form a data sample, wherein each data sample contains one of the cardiovascular and edge noise enhancement data and one of the conventional-dose CTA images.

[0107] To further improve the effectiveness and reliability of obtaining cardiovascular and edge noise enhancement data for each low-dose CTA image based on its corresponding cardiovascular and edge binary masks, a low-dose CTA image denoising model training method is provided in this application embodiment. (See also...) Figure 3 Step 130 in the low-dose CTA image denoising model training method specifically includes the following:

[0108] Step 131: Merge the cardiovascular binary mask and edge binary mask corresponding to each low-dose CTA image to obtain the key structure mask corresponding to each low-dose CTA image.

[0109] Specifically, for the cardiovascular binary mask Sv and edge binary mask S e The key structure mask obtained by merging is:

[0110] S = S v ∪S e

[0111] Step 132: Blur the region where the key structure mask is located in each of the low-dose CTA images, and add random noise to each of the low-dose CTA images to obtain the cardiovascular and edge noise enhancement data corresponding to each of the low-dose CTA images.

[0112] Specifically, the regions containing the key structure masks in each of the low-dose CTA images are blurred. The blurring process can be expressed as follows:

[0113]

[0114] Where B(·) represents the blurring operator, and x is the pixel position in the low-dose CTA image domain A; I blur (x) represents a blurred low-dose CTA image.

[0115] Then, the blurred low-dose CTA image I is processed with a preset probability p. blur (x) Add random noise N. The noise type is one of Gaussian noise, Poisson noise, or salt-and-pepper noise. The process of adding noise can be represented as:

[0116]

[0117] Among them, I aug This represents the low-dose CTA image after noise has been added, also known as cardiovascular and edge noise enhancement data Aug.A; N represents random noise, which is one of Gaussian noise, Poisson noise, or salt-and-pepper noise.

[0118] Specifically, Gaussian noise is:

[0119] I g (x)=I blur (x)+N(x)

[0120]

[0121] Among them, I g (x) represents a low-dose CTA image after adding Gaussian noise; N(x) represents Gaussian noise; σ is the standard deviation; This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise.

[0122] Poisson noise is:

[0123] I P (x)~Poission(x)

[0124] Among them, I P (x) represents a low-dose CTA image with added Poisson noise; Poisson(x) represents Poisson noise.

[0125] Salt and pepper noise is:

[0126]

[0127] Among them, I S (x) represents a low-dose CTA image with added salt-and-pepper noise; r(x) is a random number uniformly distributed in the interval [0,1], and k1 and k2 are constants.

[0128] To further improve the effectiveness and reliability of training a low-dose CTA image denoising model, in a low-dose CTA image denoising model training method provided in this application embodiment, the recurrent generative adversarial network includes: a first generative adversarial model and a second generative adversarial model;

[0129] See Figure 4 The first generative adversarial model specifically includes: a first discriminator 5, a first generator 3, and a second generator 4, wherein the output of the first generator 3 is respectively connected to the input of the second generator 4, the first discriminator 5, the cardiovascular segmentation model 1, and the edge extraction model 2;

[0130] See Figure 5 The second generative adversarial model specifically includes: a second discriminator 6, a second generator 4, and a first generator 3, wherein the output of the second generator 4 is connected to the inputs of the first generator 3, the second discriminator 6, the cardiovascular segmentation model 1, and the edge extraction model 2, respectively.

[0131] Correspondingly, see Figure 2 Step 200 in the low-dose CTA image denoising model training method specifically includes the following:

[0132] Step 210: Use the data samples to perform at least one round of iterative training on the first generative adversarial model and the second generative adversarial model. In each iteration round, based on the target loss corresponding to the first generative adversarial model, perform a preset first training step on the first generative adversarial model, and based on the target loss corresponding to the second generative adversarial model, perform a preset second training step on the second generative adversarial model.

[0133] Step 220: The first generative adversarial model after iterative training is determined as a low-dose CTA image denoising model.

[0134] To further combine vascular and edge enhancement with global denoising of low-dose CTA images during the training of a low-dose CTA image denoising model, enabling the model to balance the preservation of fine image structures with the improvement of overall image quality, in a low-dose CTA image denoising model training method provided in this application embodiment, the cardiovascular preservation loss includes: a first cardiovascular preservation loss corresponding to the first generative adversarial model; the edge preservation loss includes: a first edge preservation loss corresponding to the first generative adversarial model;

[0135] The target loss corresponding to the first generative adversarial model also includes: first identity preservation loss, first adversarial loss, and first cycle consistency loss.

[0136] Correspondingly, see Figure 3 , Figure 4 and Figure 6 The step 210 in the low-dose CTA image denoising model training method, which involves performing a preset first training step on the first generative adversarial model based on the target loss corresponding to the first generative adversarial model, specifically includes the following:

[0137] Step 211: Input the cardiovascular and edge noise enhancement data into the first generator so that the first generator outputs a first generated image corresponding to the low-dose CTA image, and obtain the first identity preservation loss between the first generated image and the cardiovascular and edge noise enhancement data.

[0138] Wherein, the first generator can be G AB Identification is performed. The first discriminator can use D. B Mark it.

[0139] Step 212: Input the first generated image into the first discriminator, the second generator, the cardiovascular segmentation model, and the edge extraction model respectively, so that the first discriminator outputs the first adversarial loss between the first generated image and the preset conventional dose CTA image, the second generator outputs the noise reduction result data corresponding to the low dose CTA image, the cardiovascular segmentation model outputs the first cardiovascular preservation loss between the first generated image and the cardiovascular binary mask, and the edge extraction model outputs the first edge preservation loss between the first generated image and the edge binary mask.

[0140] Step 213: Obtain the first cycle consistency loss between the cardiovascular and edge noise enhancement data and the noise reduction result data.

[0141] Step 214: Optimize the first generator based on the first identity preservation loss, the first adversarial loss, the first cardiovascular preservation loss, the first edge preservation loss, and the first cycle consistency loss; and optimize the second generator based on the first cycle consistency loss.

[0142] To further combine vessel and edge enhancement with global denoising of low-dose CTA images during the training of the low-dose CTA image denoising model, enabling the model to balance the preservation of fine image structures with the improvement of overall image quality, in a low-dose CTA image denoising model training method provided in this application embodiment, the cardiovascular preservation loss further includes: a second cardiovascular preservation loss corresponding to the second generative adversarial model; the edge preservation loss further includes: a second edge preservation loss corresponding to the second generative adversarial model;

[0143] The target loss corresponding to the second generative adversarial model also includes: second identity preservation loss, second adversarial loss, and second cycle consistency loss;

[0144] Correspondingly, see Figure 3 , Figure 5 and Figure 7 The step 210 in the low-dose CTA image denoising model training method, which involves performing a preset second training step on the second generative adversarial model based on the target loss corresponding to the second generative adversarial model, specifically includes the following:

[0145] Step 215: Input the conventional dose CTA image into the second generator so that the second generator outputs a second generated image corresponding to the conventional dose CTA image, and obtain the second identity preservation loss between the second generated image and the conventional dose CTA image.

[0146] Step 216: Input the second generated image into the second discriminator and the second generator respectively, so that the second discriminator outputs the second adversarial loss between the second generated image and the low-dose CTA image, and the second generator outputs the reconstruction result data corresponding to the conventional dose CTA image.

[0147] The second generated image can be identified as Gen.A, and the reconstructed result data can be identified as Rec.B; the second generator can be identified as G. BA Identification is performed. The second discriminator can use D. AIdentification is performed. Furthermore, this embodiment uses a 70×70 PatchGAN (Patch Generative Adversarial Network) as both the first and second discriminators to capture fine-grained features of the image and improve the detail quality of the generated image. Simultaneously, a gradient penalty mechanism is introduced to enhance the stability of the discriminator, preventing it from becoming too powerful during training, thereby helping the generator learn more effectively. Experimental evaluations of these improvements will be demonstrated in subsequent sections of this embodiment.

[0148] And, step 217: input the second generated image and the conventional dose CTA image into the cardiovascular segmentation model so that the cardiovascular segmentation model outputs the corresponding second cardiovascular preservation loss, and input the second generated image and the conventional dose CTA image into the edge extraction model so that the edge extraction model outputs the corresponding second edge preservation loss.

[0149] Step 218: Obtain the second cycle consistency loss between the conventional dose CTA image and the reconstructed result data.

[0150] Step 219: Optimize the second generator based on the second identity preservation loss, the second adversarial loss, the second cardiovascular preservation loss, the second edge preservation loss, and the second cycle consistency loss; and optimize the first generator based on the second cycle consistency loss.

[0151] Based on this, the target loss of this application is described in detail below:

[0152] (1) Adversarial losses (including first adversarial losses and second adversarial losses):

[0153] Adversarial loss is a fundamental component of CycleGAN, originating from the framework of Generative Adversarial Networks (GANs). Its main purpose is to incentivize the generator to produce realistic, noise-free images, deceiving the discriminator into being unable to distinguish between the generated and real images. In image denoising tasks, the generator transforms noisy images into noise-free ones, while the discriminator is responsible for differentiating between the generated noise-free images and the real noise-free images. Adversarial Loss as follows:

[0154]

[0155] in:

[0156]

[0157] in, Indicates the losses incurred in the first confrontation. Indicates the second instance of combat loss; Discriminator D BLog-likelihood expectation over the real image domain B; Discriminator D B The prediction is that the pseudo-B image generated from domain A is the log-likelihood expectation of the real B image; Discriminator D A Log-likelihood expectation over the real image domain A; Discriminator D A The prediction of generating a pseudo-A image from the B domain is the log-likelihood expectation of the real A image.

[0158] (2) Cyclic consistency loss (including the first cycle consistency loss and the second cycle consistency loss):

[0159] Relying solely on adversarial loss may lead the generator to learn arbitrary mappings, failing to guarantee the rationality and consistency of the transformation process. To address this, CycleGAN introduces cycle consistency loss, ensuring that the original image can be recovered through bidirectional transformation between the two generators. It is expressed as follows:

[0160]

[0161] in, This represents the first cycle consistency loss, which is the L1 distance between image A and the original image A after a bidirectional mapping from A to B and back to A. This represents the second cyclic consistency loss, which is the L1 distance between the image B and the original image B after a bidirectional mapping from B to A and back to B.

[0162] (3) Loss of identity preservation (including first loss of identity preservation and second loss of identity preservation):

[0163] Identity preservation loss ensures that the generator does not unnecessarily modify images already in the target domain, thus preserving their original structural and color information. The identity preservation loss... The expression is as follows:

[0164]

[0165] in, This represents the first identity preservation loss, i.e., the image in domain A is directly passed through G. BA Generate an image in domain A and calculate the L1 distance between it and the original image A. This represents the second identity preservation loss, i.e., the B-domain image is directly passed through G. AB Generate a B-domain image and calculate the L1 distance between it and the original B-domain image.

[0166] (4) Cardiovascular preservation loss (including primary and secondary cardiovascular preservation loss):

[0167] To improve the representation of cardiovascular vessels in generated CTA images for morphological evaluation, embodiments of this application introduce a cardiovascular preservation loss in Cycle GAN, designed to guide the generation of cardiovascular vessels in a targeted manner. This loss is implemented based on a pre-trained segmentation model whose parameters remain unchanged during the optimization process of Cycle GAN.

[0168] Taking the estimation of conventional dose CTA images from low-dose CTA images as an example, the original low-dose CTA image is first subjected to localized data augmentation processing. Then, the data-augmented low-dose CTA image is input into the generator G. XY A standard-dose CTA image is generated. Next, a cardiovascular segmentation model is used to segment the generated standard-dose CTA image using a cardiovascular mask. The cardiovascular binary segmentation mask of the original low-dose CTA image is used as a label to calculate the cardiovascular preservation loss. This loss prompts the generator G... XY Greater attention is being paid to the quality of cardiovascular tissue formation. Cardiovascular retention loss. The expression is:

[0169]

[0170] In the above formula, LPDE(·) represents local personalized data augmentation, f vs Represents a cardiovascular segmentation network. The first cardiovascular retention loss is represented by the L1 distance between the blood vessels segmented from the original A-domain image and the blood vessels segmented from the generated B-domain image. This represents the second cardiovascular retention loss, which is the L1 distance between the blood vessels obtained from the original B-domain image segmentation and the blood vessels obtained from the generated A-domain image segmentation.

[0171] (5) Edge preservation loss (including first edge preservation loss and second edge preservation loss):

[0172] To avoid inconsistencies in content during low-dose CTA image denoising using Cycle GAN, this embodiment incorporates an edge-preserving loss into the Cycle GAN, based on the concept of vascular-targeted enhancement. This embodiment uses a pre-trained edge extractor to extract edges from the image, constraining the edges of the generated conventional-dose CTA image to remain consistent with the edges of the original low-dose CTA image, thereby guiding the generator's learning process.

[0173] Taking the estimation of conventional-dose CTA images from low-dose CTA images as an example, the edges of the original low-dose CTA image are first extracted and used as binary labels for the subsequent generation of conventional-dose CTA images. Then, the locally personalized enhanced low-dose CTA image is input into the generator G.XY The generated conventional-dose CTA image is obtained. Edges of the generated conventional-dose CTA image are extracted using an edge extractor, and the binary edge mask of the original low-dose CTA image is used as a label to calculate the edge-preserving loss. The generator G is then optimized using this loss. XY This ensures that the edges of the generated image remain consistent with the edges of the original image, avoiding deformation. (Edge Preservation Loss) The expression is:

[0174]

[0175] In the above formula, f es This represents the edge extraction network; This represents the first edge preservation loss, which is the L1 distance between the edges obtained from the segmentation of the original A-domain image and the edges obtained from the segmentation of the generated B-domain image. This represents the second edge preservation loss, which is the L1 distance between the edges obtained from the segmentation of the original B-domain image and the edges obtained from the segmentation of the generated A-domain image.

[0176] Based on the description of the loss above, the target loss used to train the EPD-CGAN network is... for:

[0177]

[0178] Therefore, the network can be optimized through iterative alternating training:

[0179]

[0180] In order to further enable the first generator and the second generator to better preserve and enhance high-frequency details, in the low-dose CTA image denoising model training method provided in this application embodiment, both the first generator and the second generator adopt a U-Net generator with a high-frequency compensation module.

[0181] Among them, see Figure 8 The U-Net generator is pre-generated by replacing all skip connections in the U-Net network with high-frequency compensation modules (HFCModules). Here, CONV represents a convolutional layer; TCONV represents a transposed convolutional layer.

[0182] Specifically, this application proposes a High-Frequency Compensation Module (HFCModule) to enhance the generator in the EPDC-GAN framework, enabling the conversion from low-dose CTA images to conventional-dose CTA images. Specifically, this application replaces the skip connections in the U-Net generator with the HFCModule to better preserve and enhance high-frequency details, which is crucial for medical image denoising.

[0183] The HFCModule is designed to extract and utilize high-frequency components from the input image. Embodiments of this application use the Laplacian and Sobel operators to extract high-frequency information:

[0184]

[0185] Among them, K lap K represents the Laplace operator kernel; sobel_i K represents the transverse Sobel operator kernel; sobel_j This represents the vertical Sobel operator kernel.

[0186] High-frequency information is obtained by convolving the input feature x with these kernels:

[0187] HF lap =x*K lap HF sobel_i =x*K sobel_i HF sobel_j =x*K sobel_j

[0188]

[0189] HF combined =HF lap +HF sobel

[0190] Among them, HF lap This represents the result of convolving the feature map x with a Laplacian kernel; HF sobel_i This represents the result of convolving the feature map x with a horizontal Sobel operator kernel; HF sobel_j The result is obtained by convolving the feature map x with the vertical Sobel operator kernel; HF sobel The feature map x is represented by the fused result obtained by convolution with horizontal and vertical Sobel operator kernels; HF combined This represents the result of adding the feature map x through the Sobel kernel convolution and the Laplacian kernel convolution.

[0191] Next, spatial attention and channel attention mechanisms are applied to highlight important features. The fusion process is represented as follows:

[0192] Fusion = FusionModule(SA(HF) combined )+CA(x))

[0193] Where SA(·) represents spatial attention mechanism, CA(·) represents channel attention mechanism, FusionModule(·) represents feature fusion module, and Fusion represents fused features. Finally, in this embodiment, a learnable parameter α is used to control the adaptive residual connection, adding the original input feature x to the fused features:

[0194] Output = x + α·Fusion

[0195] This design enables the network to adaptively balance original features and high-frequency compensated features, enhancing its ability to recover subtle structures while effectively suppressing noise and improving the quality of the denoised image. By explicitly extracting and compensating for high-frequency components and combining them with an attention mechanism, the HFCModule significantly enhances the generator's ability to preserve important details in medical images.

[0196] In other words, the EPD-CGAN provided in this application embodiment is based on cycle GAN and utilizes two generators G AB and G BA The mapping from low-dose CTA images to conventional-dose CTA images and from conventional-dose CTA images to low-dose CTA images are learned separately, while two discriminators D are used simultaneously. A and D B The system separately determines the realism of images in the conventional-dose CTA image domain and the low-dose CTA image domain, where A represents a low-dose CTA image and B represents a conventional-dose CTA image. A game-like interaction between the generator and the discriminator forces the generator to synthesize more realistic conventional-dose and low-dose CTA images. Finally, the generator G... AB Low-dose CTA images are translated into conventional-dose CTA images to achieve image denoising. However, the cycle consistency loss used by Cycle GAN only ensures that the generated image can reconstruct the original image after inverse transformation, without explicitly constraining the synthesized image to maintain structural consistency with the original image. Furthermore, this global denoising may neglect local detail enhancement for fine structures such as blood vessels. Therefore, embodiments of this application introduce edge and blood vessel preservation losses between the synthesized image and the original input image to guide the generator training process.

[0197] Furthermore, in one example, a gradient penalty mechanism can be used to stabilize the network training process, and a self-attention CTB module can be combined to capture long-range dependencies. This application proposes a personalized image data augmentation strategy that locally enhances blood vessels and edges in images, combining this with global noise reduction processing to enhance the model's ability to enhance blood vessels and reduce noise. Specific details are as follows:

[0198] In another example of the first and second generators provided in this application embodiment, to avoid the chessboard effect, the Unet (U-shaped network) architecture can be adopted. The original 3x3 convolutions and transposed convolutions are replaced with 4x4 convolutions or transposed convolutions, with the stride of all convolutions set to 2, and the network depth is increased to improve performance. Based on this, this application embodiment replaces the skip connections in the original Unet with self-attention modules to capture long-distance dependencies.

[0199] In one example, in traditional convolutional neural networks, the convolution operation has a local receptive field, capturing only spatial features within a limited range. This limitation makes it difficult for the model to effectively model the correlation between distant pixels in an image, especially when noise has global correlation. Fixed-size convolutional kernels also struggle to adapt to the diversity and complexity of noise patterns, resulting in limited denoising performance. To address these issues, this example replaces the original skip connections in the first and second generators of the Unet structure with self-attention modules. The self-attention mechanism can capture the correlation between distant locations in the feature map, effectively modeling global information. By introducing a self-attention module between the encoder and decoder, the model can dynamically focus on important feature regions during the decoding stage, suppressing noise interference in image reconstruction.

[0200] In one example, gradient penalty is a regularization method used in generative adversarial networks (GANs) to address instability issues during network training, particularly the vanishing or exploding gradient phenomena. Its core idea is to constrain the gradient of the discriminator to satisfy the Lipschitz continuity condition, ensuring the gradient remains within a reasonable range and thus improving training stability. Specifically, implementing gradient penalty typically involves modifying the model's loss function by adding a penalty term proportional to the second norm of the gradient (i.e., the square root of the sum of squared gradients). This penalty term can be the square of the difference between the gradient norm of the first and second discriminators at the interpolation points between real and generated samples and 1, encouraging the discriminator's gradient norm at the interpolated samples to approach 1. This satisfies the 1-Lipschitz continuity condition, which helps improve the training stability of GANs.

[0201] In one example, the U-shaped generator structure of CycleGAN can be improved by replacing the traditional skip connections with a spatial and channel "squeeze and excite" (SCSE) module. The SCSE module enhances the network's ability to perceive key features by recalibrating the feature responses through a fusion of spatial and channel-dimensional attention mechanisms.

[0202] Specifically, the SCSE module consists of two parallel sub-modules: Channel Squeezing and Activation (cSE) and Spatial Squeezing and Activation (SSE). The CSE module captures the correlation between channels, generates channel weights using global average pooling and fully connected layers, and adaptively adjusts the contribution of each channel. The SSE module focuses on the spatial dimension, generating a spatial weight map through 1×1 convolutions to emphasize important spatial locations in the feature map.

[0203] In image denoising tasks, noise typically interferes with the true image signal in a high-frequency or random manner. By integrating the SCSE module, the network can adaptively enhance the response to useful features while suppressing irrelevant or redundant information introduced by noise. This mechanism improves the discriminative power of feature representations, enabling the network to more effectively separate signal from noise, thereby enhancing the denoising effect.

[0204] Furthermore, replacing traditional skip connections with SCSE modules avoids directly copying low-level features, reducing noise propagation in the network. In contrast, SCSE modules selectively deliver important information through an attention mechanism, further enhancing the network's noise reduction capabilities and learning efficiency.

[0205] Based on the above embodiments of the low-dose CTA image denoising model training method, this application also provides embodiments of the low-dose CTA image denoising method, see [link to embodiments]. Figure 9 The low-dose CTA image denoising method specifically includes the following:

[0206] Step 300: Input the target low-dose CTA image into a preset low-dose CTA image denoising model so that the low-dose CTA image denoising model outputs the denoising result data corresponding to the target low-dose CTA image, wherein the low-dose CTA image denoising model is pre-trained based on the low-dose CTA image denoising model training method.

[0207] It is understood that the low-dose CTA image denoising model training method mentioned in the embodiments of the low-dose CTA image denoising method provided in this application can be implemented based on the processing flow of the embodiments of the low-dose CTA image denoising model training method in the above embodiments. The details will not be repeated here, but can be referred to the detailed description of the embodiments of the low-dose CTA image denoising model training method described above.

[0208] Step 400: Output the noise reduction result data.

[0209] As can be seen from the above description, the low-dose CTA image denoising method provided in this application embodiment can effectively improve the application effectiveness and reliability of the low-dose CTA image denoising results obtained by using the model, thereby improving the reliability of CTA clinical diagnosis.

[0210] From a software perspective, this application also provides a low-dose CTA image denoising model training apparatus for performing all or part of the low-dose CTA image denoising model training method, see [link to relevant documentation]. Figure 10 The low-dose CTA image denoising model training device specifically includes the following components:

[0211] Data augmentation module 10 is used to acquire cardiovascular and edge noise enhancement data corresponding to low-dose CTA images based on cardiovascular segmentation model and edge extraction model;

[0212] The model training module 20 is used to train a preset recurrent generative adversarial network based on the cardiovascular and edge noise enhancement data and the preset conventional dose CTA image, so as to train the recurrent generative adversarial network into a low dose CTA image denoising model. The target loss includes cardiovascular preservation loss obtained based on the cardiovascular segmentation model and edge preservation loss obtained based on the edge extraction model.

[0213] The embodiments of the low-dose CTA image denoising model training device provided in this application can be used to execute the processing flow of the embodiments of the low-dose CTA image denoising model training method described above. Its functions will not be repeated here, but can be referred to the detailed description of the embodiments of the low-dose CTA image denoising model training method described above.

[0214] The training portion of the low-dose CTA image denoising model in the aforementioned low-dose CTA image denoising model training device can be performed on a server or a client device. The choice can be made based on the processing capabilities of the client device and limitations of the user's usage scenario. This application does not impose any limitations in this regard. If all operations are performed on the client device, the client device may further include a processor for specific processing during the low-dose CTA image denoising model training.

[0215] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0216] The server and the client device can communicate using any suitable network protocol, including those not yet developed as of the date of this application. Such network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Furthermore, such network protocols may also include RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer Protocol) protocols used on top of the aforementioned protocols.

[0217] As can be seen from the above description, the low-dose CTA image denoising model training device provided in this application embodiment can combine blood vessel and edge enhancement with global denoising of low-dose CTA images during the low-dose CTA image denoising model training process, so that the model can take into account both the preservation of the fine structure of the image and the improvement of the overall image quality, thereby effectively improving the application effectiveness and reliability of the low-dose CTA image denoising results obtained by using the model, so as to improve the reliability of CTA clinical diagnosis.

[0218] From a software perspective, this application also provides an apparatus for performing all or part of the low-dose CTA image denoising method described above, see [link to relevant documentation]. Figure 11 The low-dose CTA image noise reduction device specifically includes the following components:

[0219] The image denoising module 30 is used to input the target low-dose CTA image into a preset low-dose CTA image denoising model, so that the low-dose CTA image denoising model outputs the denoising result data corresponding to the target low-dose CTA image, wherein the low-dose CTA image denoising model is pre-trained based on the low-dose CTA image denoising model training method.

[0220] The data output module 40 is used to output the noise reduction result data.

[0221] The embodiments of the low-dose CTA image denoising device provided in this application can be used to execute the processing flow of the embodiments of the low-dose CTA image denoising method described above. Its functions will not be repeated here, but can be referred to the detailed description of the embodiments of the low-dose CTA image denoising method described above.

[0222] The low-dose CTA image denoising device can perform the low-dose CTA image denoising portion in either a server or a client device. The choice depends on the processing power of the client device and limitations imposed by the user's usage scenario. This application does not impose any limitations in this regard. If all operations are performed in the client device, the client device may further include a processor for the specific processing of the low-dose CTA image denoising.

[0223] As can be seen from the above description, the low-dose CTA image denoising device provided in this application embodiment can effectively improve the application effectiveness and reliability of the low-dose CTA image denoising results obtained by using the model, thereby improving the reliability of CTA clinical diagnosis.

[0224] To further illustrate the performance of the low-dose CTA image denoising model provided in this application, specifically to quantitatively evaluate the performance of EPD-CGAN, given that the training data is unpaired, the verification examples in this application use two image quality metrics to evaluate the quality of the generated CTA images: BRISQUE (Blind / Referenceless Image Spatial Quality Evaluator) and the Tenengrad gradient function. BRISQUE is a classic no-reference image quality evaluation method. Unlike traditional reference-based evaluation methods (such as PSNR and SSIM), BRISQUE does not require an original noise-free image as a control; instead, it quantifies the quality by analyzing the statistical characteristics of the image itself. The core idea is to use the statistical characteristics of natural scenes to describe the local structure and texture distribution of the image, and then model these features using a support vector machine to predict a quantitative quality score for a given image. The lower the score, the closer the image is to a natural, high-quality state; the higher the score, the more significant the distortion in the image, such as noise, blur, or artifacts.

[0225] The Tenengrad gradient function is a no-reference sharpness evaluation metric based on image gradient information. Its basic idea is to measure image sharpness and clarity by detecting the intensity of high-frequency components such as edges and textures. Specifically, the Tenengrad method typically first uses the Sobel operator to extract gradients in the horizontal and vertical directions of the image, and then calculates the statistics of the gradient magnitudes to obtain a sharpness score. (See below.)

[0226]

[0227] Among them, G h (i,j) and G h (i,j) represent the horizontal and vertical gradient values ​​of the image at pixel (i,j), respectively, calculated by the Sobel operator. Furthermore, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were used as evaluation metrics on the Mayo dataset to validate the validity of BRISQUE and Tenengrad scores in measuring image quality.

[0228] The experimental environment used the PyTorch framework, version 1.12.1. During both training and testing, all experiments were performed on an NVIDIA A6000 graphics card and an Intel i9-12900K CPU. When training the dataset, each 512×512 low-dose CTA image was scaled to 286×286 and then randomly cropped to 256×256 for data augmentation; the corresponding regular-dose CTA images were also augmented in the same way. Furthermore, the validation example performed personalized data augmentation on the low-dose CTA images, blurring blood vessels or edges, and randomly adding noise to the entire image with a 0.3 chance. The noise type was one of Gaussian noise, Poisson noise, or salt-and-pepper noise. To ensure model convergence, the validation example ran for 1000 training epochs, with the weights of adversarial loss, cycle consistency loss, identity loss, vessel preservation loss, and edge preservation loss set to 1, 10, 5, 1, and 1, respectively. During training, the network parameters are optimized by minimizing the composite loss function using the Adam optimizer, and the learning rate is increased from 2×10⁻⁶ using cosine annealing. -4 Gradually decrease to 1×10 -6 During the inference process, this application verifies that low-dose CTA images of 512×512 are cropped into overlapping patches to evaluate the network's performance.

[0229] This application's validation examples demonstrate the denoising capabilities of the model in real low-dose CTA images through extensive experiments. It should be noted that, since the validation examples are trained on unpaired data, they will be compared with unsupervised denoising models, including BM3D, Noise2Noise, Noise2Void, noise2sim, and IDR.

[0230] 1) CTP2CTA Dataset: Validation examples in this application evaluate the performance of the proposed EPD-CGAN model against various benchmark denoising methods on the CTP2CTA test set, including the traditional BM3D method and unsupervised denoising frameworks such as Noise2Noise, Noise2Void, Noise2Sim, and IDR. In this experiment, CTA images are treated as regular-dose CTA images, while CTP images are treated as low-dose CTA images.

[0231] Since the validation examples in this application were trained using unpaired low-dose CTA images and conventional-dose CTA image data, traditional evaluation metrics requiring reference standards were not applicable. Therefore, referenceless image quality metrics such as BRISQUE and Tenengrad were adopted as objective quantitative standards. These metrics corroborate the qualitative observation results, providing strong evidence for the advantages of the model in detail preservation and structural enhancement in the validation examples of this application.

[0232] Experiments demonstrate that, compared to classic denoising schemes such as BM3D and Noise2Noise, the model proposed in the validation examples of this application not only effectively suppresses random noise in low-dose CTA images but also demonstrates superior performance in preserving and even enhancing high-frequency details. BM3D and Noise2Noise often sacrifice some texture and edge information during noise reduction; that is, the final generated images often exhibit slight oversmoothing, resulting in a loss of clarity in blood vessel contours and organ edges while they become cleaner. Although Noise2Void, Noise2Sim, IDR, and the original CycleGAN model do not over-blur high-frequency features, they are still insufficient in enhancing tissue details and highlighting key structures. In other words, while preserving the basic texture of the image, these methods do not make organ edges sharper or blood vessels more clearly discernible. In comparison, the verification example model of this application combines noise reduction with high-frequency detail preservation by incorporating vessel preservation loss and edge preservation loss, and integrating a high-frequency compensation module in the generator. This reduces noise while more clearly highlighting the differences in blood vessel, organ contours and tissue structures, and visually demonstrates a clear advantage with clearer blood vessel morphology and more prominent organ edges.

[0233] Based on the analysis results in Table 1, the validation example model of this application significantly outperforms other unsupervised methods, both in terms of the quantitative improvement of the BRISQUE metric and the measurement of image gradient and sharpness by the Tenengrad metric. From a quantitative analysis perspective, the proposed model improves the performance of low-dose CTA images on both the BRISQUE and Tenengrad unreferenced metrics to a level comparable to that of conventional-dose CTA images in the test set.

[0234] In summary, the unsupervised low-dose CTA image denoising model proposed in this application demonstrates significant advantages over various benchmarks and existing methods, resulting in clearer vascular morphology and more prominent organ edges. Supported by both reference-free quantitative metrics and visual observation, this model effectively balances denoising with structural enhancement.

[0235] Table 1

[0236]

[0237]

[0238] 2) Mayo Dataset: To compensate for the limitations of the lack of reference metrics in evaluating this dataset, this application's validation example introduces additional reference-based evaluation metrics PSNR and SSIM on the Mayo dataset, thereby verifying the rationality of BRISQUE and Tenengrad scores in measuring image quality. Since vascular structures are difficult to observe clearly in the Mayo dataset, this application's validation example does not use vessel preservation loss in this experiment, but still retains the high-frequency compensation module and edge preservation loss. In this dataset, the noise in low-dose CTA images is not as significant as in CTP2CTA, making it difficult to highlight the performance advantages of traditional or self-supervised denoising methods. From a visual perspective, qualitative analysis shows that while BM3D, Noise2Noise, Noise2Void, and Noise2Sim can reduce some noise, they often lead to the loss of high-frequency details, ultimately failing to significantly improve the structural clarity of the images. IDR also shows insufficient denoising capability in such low-noise scenarios, and the original CycleGAN method also fails to enhance edge features, failing to achieve higher quality restoration results compared to conventional dose CTA images. In contrast, the model proposed in the verification examples of this application can effectively remove image noise while still fully preserving the structural and textural features of the image, thus achieving the same level of visual effect as the reference image.

[0239] Furthermore, according to Table 2, the quantitative analysis results show that the model proposed in the validation examples of this application achieved optimal performance in key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Blind / Referenceless Image Spatial Quality Evaluator (BRISQUE). Simultaneously, this model improved the quality level of low-dose CTA images on the BRISQUE and Tenengrad metrics to a level approaching that of conventional-dose CTA images. It is worth noting that compared to this model, methods such as BM3D, Noise2Noise, Noise2Void, Noise2Sim, and IDR significantly lowered the Tenengrad metric due to excessive smoothing processing, thus losing valuable texture detail. Meanwhile, the trends in PSNR and SSIM results show that the performance improvements demonstrated by these supervised metrics are consistent with the improvements shown by BRISQUE and Tenengrad, further confirming the effectiveness and rationality of BRISQUE and Tenengrad in measuring image quality.

[0240] Table 2

[0241]

[0242]

[0243] In addition, this application designed several ablation studies to analyze and verify the performance of the model in this application under different model structures and composite loss configurations:

[0244] 1) Skip Connection Variants: To verify the effectiveness of different skip connection modules in low-dose CT (low-dose CTA image) denoising, this application's validation example replaced the skip connections in the U-shaped architecture of the s-generator with various modules. Specifically, the original skip connections were replaced with the scSE module, the self-attention module, the CBAM (Convolutional Block Attention Module), and the high-frequency compensation module proposed in this application's validation example. The performance of each module on the test set was compared to evaluate its impact on the overall model performance. Experimental results are shown in Table 3, and the performance of each module on BRISQUE and Tenengrad metrics is as follows:

[0245] Table 3

[0246] Module type BRISQUE Tenengrad Low-dose CTA images 30.4254± 45.4398± Original Unet jump connection 23.7063±7.5041 53.2979±10.7501 scSE module 26.4185±7.9227 51.6354±10.0581 Self-attention module 24.0141±6.3786 50.1862±8.1440 CBAM module 23.1034±8.7140 53.3586±9.1266 High frequency compensation module 22.8622±7.4361 54.1680±10.5241

[0247] As shown in Table 3, although most replacement schemes (including scSE and the self-attention module) failed to effectively improve the model's performance on the BRISQUE and Tenengrad no-reference metrics compared to the original identity skip connections, CBAM and the high-frequency compensation module showed some improvement. The high-frequency compensation module performed particularly well, surpassing not only the original skip connections and other modules in preserving high-frequency details and enhancing edge sharpness, but also CBAM. In other words, the high-frequency compensation module has a stronger ability than CBAM to simultaneously reduce noise and maintain key microstructural information, providing stronger experimental evidence for the performance advantage of the framework proposed in this application's validation examples on unpaired low-dose CTA image data.

[0248] 2) Composite Loss: To explore the impact of different loss function combinations on the denoising performance of low-dose CT (low-dose CTA images), this application's verification example extends the original CycleGAN loss function. The original CycleGAN loss function consists of three parts: adversarial loss, cycle consistency loss, and identity loss. To better preserve the contours of blood vessels and edges during denoising and prevent image distortion, this application's verification example introduces edge preservation loss and blood vessel preservation loss. By comparing the performance of different loss function combinations on the test set, the improvement effect of the added loss function on model performance is evaluated. All models are trained and evaluated on the same training and test sets to ensure the fairness of the experimental results. This application's verification example designs the following loss function combinations for comparative experiments:

[0249] (1) Original CycleGAN loss: including adversarial loss, cycle consistency loss and identity loss.

[0250] (2) Add edge preservation loss: Add edge preservation loss to the original CycleGAN loss.

[0251] (3) Adding vascular retention loss: In addition to the original CycleGAN loss, vascular retention loss is added.

[0252] (4) Simultaneous addition of edge preservation loss and vessel preservation loss: On the basis of the original CycleGAN loss, edge preservation loss and vessel preservation loss are added simultaneously.

[0253] The experimental results are shown in Table 4. The performance of different combinations of loss functions on PSNR and SSIM metrics is as follows:

[0254] Table 4

[0255]

[0256] As shown in Table 4, compared to using only the original loss, both adding edge preservation loss and vessel preservation loss improve the model's denoising performance without reference metrics. Furthermore, from a metric comparison perspective, there is no significant difference in quantitative scores between the model with the original loss plus edge preservation loss and the model with the original loss plus vessel preservation loss. However, in visual comparison, the two models emphasize different aspects of improvement. The model with edge preservation loss exhibits a sharper rendering effect on image contours and structural edges, making boundaries clearly distinguishable and avoiding the problem of over-smoothing; while the model with vessel preservation loss excels in preserving fine vascular structures, making vessels highly recognizable and resolving in the denoised image. When these two losses are used simultaneously, the model achieves a synergistic improvement in overall denoising quality, edge sharpness, and vascular detail rendering, providing a more comprehensive guarantee for the final image reconstruction quality and clinical usability.

[0257] 3) Personalized Data Augmentation: To investigate the impact of personalized data augmentation strategies on the denoising performance of low-dose CT (low-dose CTA images), this application's verification example performed specific augmentation processing on the training data. Specifically, this application's verification example blurred the blood vessels and edges in the low-dose CTA images, and added Gaussian noise, Poisson noise, and salt-and-pepper noise to the entire image. By comparing the performance of different data augmentation strategies on the test set, their improvement effect on model performance was evaluated. This application's verification example designed the following data augmentation strategies for ablation experiments:

[0258] (1) Training was performed using the original low-dose CTA images without any data augmentation.

[0259] (2) Gaussian blurring is applied to blood vessels and edge parts in low-dose CTA images to simulate the loss of image details and enhance the model’s sensitivity to detail information.

[0260] (3) Add Gaussian noise to low-dose CTA images.

[0261] (4) Add Poisson noise to low-dose CTA images.

[0262] (5) Add salt and pepper noise to low-dose CTA images.

[0263] (6) Comprehensive data augmentation strategy: Apply all the above data augmentation methods at the same time to maximize the diversity of training data.

[0264] The experimental results are shown in Table 5. The performance of different data augmentation strategies on PSNR and SSIM metrics is as follows:

[0265] Table 5

[0266] Data augmentation strategy BRISQUE Tenengrad Low-dose CTA images 30.4254± 45.4398± No data augmentation 25.9490±5.9117 52.4128±9.1454 blurring of blood vessels and edges 23.6930±8.0475 53.9606±10.6967 Add Gaussian noise 22.5430±8.3617 50.7760±10.0710 Adding Poisson noise 24.4023±7.3253 52.4311±9.110 Add salt and pepper noise 24.6689±6.1163 52.5255±9.7299 Comprehensive data augmentation strategy 22.8622±7.4361 54.1680±10.5241

[0267] As shown in Table 5, all these methods improved the BRISQUE score, with the Gaussian noise addition method showing the most significant improvement. However, this strategy resulted in a decrease in the Tenengrad score. This contrasts sharply with the results of adding blood vessel and edge blurring: although the improvement in BRISQUE was relatively limited, it significantly improved Tenengrad. This phenomenon indicates that personalized blood vessel and edge blurring enhancement strategies can effectively unleash the model's potential in fine structure and edge sharpening, thus achieving better performance in high-frequency detail preservation and sharpening.

[0268] In other words, this application's verification example proposes using Cycle GAN to translate low-dose CTA images into conventional-dose CTA images. To this end, this application's verification example collected a dataset of low-signal-to-noise ratio (SNR) low-dose CTA images with blurred blood vessels and a dataset of high SNR images with clear blood vessels to train the model. For these two datasets, this application's verification example improved the model's generator to enhance its generation capability. Simultaneously, this application's verification example proposes local data augmentation strategies targeting blood vessels and image edges. Furthermore, to make the generator focus more on blood vessel enhancement, a blood vessel extraction network is introduced to guide the generator's optimization, and an edge extraction network is also introduced to improve the generator's texture preservation capability. To the best of our knowledge, this application's verification example is the first to propose an innovative method combining local blood vessel enhancement with global image denoising, filling the gap in existing technologies in simultaneously preserving fine structures and improving overall image quality. The main contributions of this application's verification example can be summarized as follows:

[0269] (1) The verification example of this application collected data of 70 low-dose CTA images and 27 conventional-dose CTA images for training EPD-CGAN to achieve image denoising and vascular enhancement of low-dose CTA images.

[0270] (2) The verification examples of this application propose local data enhancement techniques for blood vessels and edges in CTA images to increase data diversity.

[0271] (3) The verification example of this application proposes a method of jointly guiding optimization using a blood vessel extraction network and an edge extraction network, which makes the model pay more attention to the generation of blood vessels and edges while reducing noise.

[0272] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the low-dose CTA image denoising model training method and / or the low-dose CTA image denoising method mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and the memory via wired or wireless means.

[0273] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0274] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the low-dose CTA image denoising model training method and / or the low-dose CTA image denoising method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the low-dose CTA image denoising model training method and / or the low-dose CTA image denoising method in the above method embodiments.

[0275] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0276] The one or more modules are stored in the memory, and when executed by the processor, they execute the low-dose CTA image denoising model training method and / or the low-dose CTA image denoising method in the embodiment.

[0277] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0278] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0279] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0280] This application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the aforementioned low-dose CTA image denoising model training method and / or low-dose CTA image denoising method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0281] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned low-dose CTA image denoising model training method and / or low-dose CTA image denoising method.

[0282] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0283] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0284] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0285] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a low-dose CTA image denoising model, characterized in that, The application relates to a method for training a low-dose CTA image denoising model. The method comprises the following steps: obtaining cardiovascular and edge noise enhancement data corresponding to a low-dose CTA image based on a cardiovascular segmentation model and an edge extraction model; the cardiovascular and edge noise enhancement data are generated by performing fuzzy processing on a region where a key structure mask corresponding to a cardiovascular binary mask and an edge binary mask is located; training a preset cycle generative adversarial network based on a target loss to train the cycle generative adversarial network into a low-dose CTA image denoising model according to the cardiovascular and edge noise enhancement data and a preset regular-dose CTA image, wherein the target loss comprises a cardiovascular preservation loss obtained based on the cardiovascular segmentation model and an edge preservation loss obtained based on the edge extraction model; the low-dose CTA image corresponds to a radiation dose smaller than that of the regular-dose CTA image; the cycle generative adversarial network comprises a first generative adversarial model and a second generative adversarial model; the method comprises the following steps: performing at least one round of iterative training on the first generative adversarial model and the second generative adversarial model based on a data sample, and in each iteration round, performing a preset first training step on the first generative adversarial model based on a target loss corresponding to the first generative adversarial model, and performing a preset second training step on the second generative adversarial model based on a target loss corresponding to the second generative adversarial model; determining the first generative adversarial model after the iterative training as a low-dose CTA image denoising model; the method comprises the following steps: inputting the cardiovascular and edge noise enhancement data into a first generator to enable the first generator to output a first generated image corresponding to the low-dose CTA image, and obtaining a first identity preservation loss between the first generated image and the cardiovascular and edge noise enhancement data; inputting the first generated image into a first discriminator, a second generator, the cardiovascular segmentation model and the edge extraction model respectively to enable the first discriminator to output a first adversarial loss between the first generated image and a preset regular-dose CTA image, enable the second generator to output denoising result data corresponding to the low-dose CTA image, enable the cardiovascular segmentation model to output a first cardiovascular preservation loss between the first generated image and a cardiovascular binary mask, and enable the edge extraction model to output a first edge preservation loss between the first generated image and an edge binary mask; obtaining a first cycle consistency loss between the cardiovascular and edge noise enhancement data and the denoising result data; optimizing the first generator based on the first identity preservation loss, the first adversarial loss, the first cardiovascular preservation loss, the first edge preservation loss and the first cycle consistency loss, and optimizing the second generator based on the first cycle consistency loss.

2. The low-dose CTA image denoising model training method of claim 1, characterized in that, The cardiovascular segmentation model and the edge extraction model are used to obtain cardiovascular and edge noise enhancement data corresponding to the low-dose CTA image, and the method comprises the following steps: obtaining a plurality of data pairs, wherein each data pair comprises a low-dose CTA image and a regular-dose CTA image; extracting a cardiovascular binary mask corresponding to each low-dose CTA image based on the cardiovascular segmentation model, and extracting an edge binary mask corresponding to each low-dose CTA image based on the edge extraction model; obtaining cardiovascular and edge noise enhancement data corresponding to each low-dose CTA image based on the cardiovascular binary mask and the edge binary mask corresponding to each low-dose CTA image; combining the cardiovascular and edge noise enhancement data corresponding to each low-dose CTA image with the regular-dose CTA image in the data pair in which the low-dose CTA image is located to obtain a data sample, wherein each data sample comprises the cardiovascular and edge noise enhancement data and the regular-dose CTA image.

3. The low-dose CTA image denoising model training method of claim 2, characterized in that, The cardiovascular segmentation model and the edge extraction model are used to obtain cardiovascular and edge noise enhancement data corresponding to the low-dose CTA image, and the method comprises the following steps: combining the cardiovascular binary mask and the edge binary mask corresponding to each low-dose CTA image to obtain a key structure mask corresponding to each low-dose CTA image; obtaining cardiovascular and edge noise enhancement data corresponding to each low-dose CTA image by performing blur processing on the region of the key structure mask in each low-dose CTA image and adding random noise to each low-dose CTA image.

4. The low-dose CTA image denoising model training method according to claim 2, wherein the first generative adversarial model comprises a first discriminator, a first generator and a second generator, wherein the output end of the first generator is connected to the input end of the second generator, the first discriminator, the cardiovascular segmentation model and the edge extraction model respectively; the second generative adversarial model comprises a second discriminator, the second generator and the first generator, wherein the output end of the second generator is connected to the input end of the first generator, the second discriminator, the cardiovascular segmentation model and the edge extraction model respectively.

5. The low-dose CTA image denoising model training method of claim 1, characterized in that, The cardiovascular preservation loss further comprises a second cardiovascular preservation loss corresponding to the second generative adversarial model, and the edge preservation loss further comprises a second edge preservation loss corresponding to the second generative adversarial model. The target loss corresponding to the second generative adversarial model further comprises a second identity preservation loss, a second adversarial loss and a second cycle consistency loss; Correspondingly, the second training step for the second generative adversarial model based on the target loss corresponding to the second generative adversarial model comprises: inputting the regular-dose CTA image into a second generator to make the second generator output a second generated image corresponding to the regular-dose CTA image, and obtaining a second identity preservation loss between the second generated image and the regular-dose CTA image; inputting the second generated image into a second discriminator and a second generator respectively to make the second discriminator output a second adversarial loss between the second generated image and the low-dose CTA image, and make the second generator output a reconstruction result data corresponding to the regular-dose CTA image; and inputting the second generated image and the regular-dose CTA image into the cardiovascular segmentation model to make the cardiovascular segmentation model output a corresponding second cardiovascular preservation loss, and inputting the second generated image and the regular-dose CTA image into the edge extraction model to make the edge extraction model output a corresponding second edge preservation loss; obtaining a second cycle consistency loss between the regular-dose CTA image and the reconstruction result data; optimizing the second generator based on the second identity preservation loss, the second adversarial loss, the second cardiovascular preservation loss, the second edge preservation loss and the second cycle consistency loss, and optimizing the first generator based on the second cycle consistency loss.

6. The low-dose CTA image denoising model training method of claim 4, characterized in that, The first generator and the second generator are both U-Net generators provided with high-frequency compensation modules. The U-Net generator is generated by replacing all the skip connections in the U-Net network with high-frequency compensation modules in advance.

7. A low-dose CTA image denoising method, characterized in that, The method comprises: inputting a target low-dose CTA image into a preset low-dose CTA image denoising model to make the low-dose CTA image denoising model output denoising result data corresponding to the target low-dose CTA image, wherein the low-dose CTA image denoising model is trained in advance based on the low-dose CTA image denoising model training method in any one of claims 1 to 6; outputting the denoising result data.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the low-dose CTA image denoising model training method in any one of claims 1 to 6, and / or implement the low-dose CTA image denoising method in claim 7.

9. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the low-dose CTA image denoising model training method in any one of claims 1 to 6, and / or implement the low-dose CTA image denoising method in claim 7.

Images