A carotid artery non-contrast CT three-dimensional angiography enhancement method based on mask guidance and reverse distillation

Abstract

The application aims to develop a carotid non-contrast CT (NCCT) three-dimensional angiography enhancement method based on mask guidance and reverse distillation. First, based on 100 pairs of pre-training data containing doctor's annotation, the carotid automatic segmentation network suitable for CTA and NCCT images is trained respectively. On this basis, the generation process is divided into two stages: in the first stage, the carotid blood vessel pseudo-labels for 1650 pairs of training data are generated by using the above segmentation model, the mask guidance loss and MIND structural loss are designed and integrated into 3D CycleGAN, an end-to-end three-dimensional angiography generation framework is constructed, and a good generator A (CTA NCCT) and an initial rough generator B (NCCT CTA) are obtained; in the second stage, based on the generator A, the blood vessel silhouette and the intermediate layer feature map are used as the bridge, and the generator B is further optimized by the reverse distillation strategy. The application can automatically generate high-quality three-dimensional virtual CTA images meeting the clinical requirements only by inputting NCCT.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and analysis technology, specifically relating to a method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation. Background Technology

[0002] Vascular diseases are among the leading causes of death worldwide, and early, accurate diagnosis of lesions such as stenosis, occlusion, and aneurysms in the head and neck vessels is crucial for clinical treatment decisions. Computed tomography (CTA), as the gold standard imaging technique in this field, significantly enhances the contrast between blood and surrounding tissues through intravenous injection of iodinated contrast agents, clearly displaying vascular anatomy and pathological features, providing key evidence for disease diagnosis. However, CTA's dependence on contrast agents brings significant clinical risks: some patients have contraindications such as contrast agent allergy, renal insufficiency, or multiple myeloma, making them unable to tolerate contrast agent injections; furthermore, repeated exposure to contrast agents may increase the risk of complications such as contrast-induced nephropathy, and the examination is costly and complex. In contrast, non-contrast CT (NCCT) does not require contrast agent injection, offering advantages such as faster acquisition speed, relatively lower radiation dose, and no contrast agent-related risks. However, it cannot effectively display internal vascular structures and is difficult to directly use for vascular lesion assessment. Therefore, how to directly generate high-quality CTA images from NCCT images based on artificial intelligence (AI) technology has become a research hotspot in the interdisciplinary field of medical imaging and artificial intelligence. This technology not only avoids the risks of contrast agent use and reduces radiation dose, but also simplifies the examination process and improves diagnostic efficiency, providing a safer and more convenient diagnostic solution for vascular diseases in clinical practice. It has significant clinical translational value and research significance.

[0003] From a technical perspective, the core of AI angiography is to precisely adjust the grayscale intensity of vascular regions in NCCT images using intelligent algorithms, simulating the enhancement effect of contrast agents on blood vessels, thereby generating high-quality images equivalent to clinical CTA. Currently, the technical approaches in this field are mainly divided into two categories: generative methods based on generative adversarial networks (GANs) and progressive generative methods based on diffusion models.

[0004] Extensive research has been conducted on GAN-based methods. An early work proposed a 2D CTA-GAN model, employing a network structure combining a generator, corrector, and discriminator. This model achieved the synthesis of images ranging from NCCT to neck and abdominal CTAs, achieving a PSNR of 32.07 dB and a SSIM of 0.919 on the internal test set, and a diagnostic accuracy of 86% on the external validation set, preliminarily validating the feasibility of contrast-free CTA generation. Another study proposed the VesselTransGAN framework, introducing the Transformer's self-attention mechanism into the GAN generator and designing a 2D-3D fusion strategy. It enhances local blood vessels by aligning pixel grayscale values ​​and fuses the dual-path generation results of 2D slices and 3D blocks, achieving an SSIM of 0.929 on 426 pairs of CT-CTA data, with particular focus on the generation of small blood vessels in the brain. Furthermore, some reports have utilized GANs to convert entire non-contrast chest CT scans into synthetic contrast-enhanced CT scans, improving the visualization of mediastinal lymph nodes; others have designed an aortic sensing GAN to enhance the depiction of abdominal aortic aneurysms by focusing on the aortic boundary; and one paper proposed a two-stage GAN framework to address spatial misalignment and patient-specific contrast uptake variability. However, GAN methods have inherent limitations: adversarial training is unstable, prone to pattern collapse and artifacts; the global discrimination mechanism struggles to focus on vascular regions with extremely small pixel proportions (less than 5% in the carotid artery); and most methods are based on 2D frameworks, ignoring the spatial continuity of 3D vessels, leading to slice breaks and misalignments during reconstruction.

[0005] In the past two years, new advances have emerged in diffusion-based methods. One study proposed the DiffCTA model, which applies an adversarial diffusion model to NCCT-to-CTA generation. By introducing an unsupervised contrast enhancement mask, designing a contrast enhancement attention mechanism, a contrast perception discriminator, and a mask-guided PatchNCE loss, the generator focuses on the vascular region affected by the contrast agent. At the same time, it uses large-step diffusion to reduce the number of inference steps from 1000 to 4, achieving a PSNR of 24.74 dB and an SSIM of 0.864 on the test set. Another work proposes an AEGIS framework that employs a conditional multi-view diffusion model. Instead of directly learning CTA, it uses contrast maps (CM = CTA - NCCT) as training targets. Through multi-view fusion learning, it learns 2D slice distributions from three planes—transverse, coronal, and sagittal—to perceive 3D features. A conditional angiography diffusion estimation module is designed to encode NCCT as spatial guidance cues. Finally, a lightweight autoencoder fuses the multi-view contrast maps, achieving excellent results of 37.08 dB PSNR and 0.972 on cardiac CT coronary artery and cardiac chamber enhancement tasks, with particularly significant enhancement effects on small coronary artery branches. Furthermore, existing literature uses conditional DDPM to synthesize contrast-enhanced dual-energy CT from NCCT, but it performs poorly in capturing small cerebral vascular systems. Another study proposed the SynDiff model, which uses an adversarial diffusion model for unsupervised medical image translation. The advantages of diffusion models are high generation quality and stable training, but their main problems include: high computational complexity in directly processing 3D data; time-consuming inverse sampling process (although it can be accelerated by increasing the step size or reducing the number of steps, but details may be lost); and most diffusion models are also based on 2D slicing processing, lacking explicit modeling of the continuity of three-dimensional space.

[0006] Besides the two mainstream methods mentioned above, some researchers have also attempted cascaded frameworks based on U-Net or Transformer. For example, some works have cascaded 3D nnU-Net with a generative network for aortic dissection detection; others have reported the proposal of a cycle-consistent Transformer model CyTran to mitigate misalignment artifacts. However, these methods generally lack explicit guidance on the vascular region, resulting in often blurred output images and the easy loss of small vascular branches.

[0007] In summary, although GANs and diffusion models have made significant progress in NCCT to CTA generation tasks, the accuracy and robustness of existing solutions still fall short of clinical practical requirements for the specific scenario of the head and neck carotid arteries (with thin vessels, complex backgrounds, and high requirements for 3D continuity). A comprehensive analysis of existing technologies reveals that they are fundamentally constrained by the following four core technological bottlenecks:

[0008] (1) Limitations of the generation paradigm: Existing technologies are mostly based on 2D generation frameworks (whether it is GAN or diffusion model, most of them are trained on 2D slices), ignoring the spatial continuity of 3D blood vessels and cross-slice correlation, which leads to problems such as blood vessel breakage, misalignment and inconsistency in strength between slices during 3D reconstruction. Although AEGIS has tried multi-view learning, it still fits 3D distribution through 2D slices and has failed to fundamentally establish an end-to-end 3D generation paradigm.

[0009] (2) Extremely low target proportion: The carotid artery accounts for less than 5% of the pixels in head and neck CT images, with thin branches and complex structure. Existing generative models generally use a global loss function, and background tissues (bone, muscle, fat) occupy most of the pixels. The model feature learning is prone to deviating from the core vascular region, resulting in insufficient or complete enhancement of small branches. Although DiffCTA introduces a contrast enhancement mask, its mask depends on the subtraction threshold and is difficult to adapt to the vascular morphology variations of different patients; AEGIS uses the contrast image as the target, but the contrast image still contains a large amount of weak difference noise from non-vascular areas.

[0010] (3) Data alignment bias: There is a time interval between NCCT and CTA scans (usually several minutes to tens of minutes), and slight shifts in patient breathing, heartbeat, and body position are inevitable, resulting in non-rigid deformation and registration errors in the paired data of the two modalities. Existing methods (especially pixel-level supervised GANs and diffusion models) directly use incompletely aligned paired images for training, treating the registration error as a learning target, introducing artifacts and structural distortion. Although contrastive learning methods such as PatchNCE loss can alleviate some of the misalignment problem, they have not completely solved it.

[0011] (4) Difficulty in boundary differentiation: In NCCT images, the Henle unit (HU) values ​​of blood vessels and surrounding soft tissues (such as muscles, lymph nodes, and nerves) highly overlap, have similar grayscale features, and have extremely low boundary recognition. The model struggles to accurately define the blood vessel contour, resulting in blurred blood vessel edges and adhesion to the background in the generated virtual CTA, and even diagnostic traps such as false stenosis or false aneurysm. Although the MIND structural loss can enhance structural consistency by utilizing local neighborhood descriptors, existing methods mostly apply it to global or simple regions, lacking targeted optimization for the tree-like branches of blood vessels. Summary of the Invention

[0012] Purpose of the invention: To address the four core technical bottlenecks in existing NCCT-based CTA generation methods—generation paradigm limitations, extremely low target area proportion, data alignment deviation, and difficulty in boundary differentiation—this invention proposes a carotid artery non-contrast CT three-dimensional angiography enhancement method based on mask guidance and reverse distillation. The aim is to automatically generate high-quality virtual CTA images that meet clinical diagnostic needs without the need for contrast agents.

[0013] Technical Solution: This invention proposes a method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation, specifically including the following steps:

[0014] a) Literature review: In-depth collection of literature and invention patents in related fields, including the current research status of NCCT-generated CTA, generative adversarial networks, diffusion models, and knowledge distillation.

[0015] b) Data Collection and Preprocessing: Paired CT and NCCT images were collected. The collected data were divided into pre-training data, training set, and test set based on whether they contained vascular annotations, and preprocessed. 1788 pairs of 3D NCCT and CTA image pairs were collected from Jinling Hospital affiliated with Nanjing University School of Medicine, and 138 pairs of images were selected for pixel-level annotation of vascular regions. 100 pairs of images with vascular annotations were used as pre-training data to train the carotid artery segmentation model; 1650 pairs of data without vascular annotations were used as the training set to train the generative model; the remaining 38 pairs of data were used as an independent test set and underwent uniform preprocessing (including resampling to the same spatial resolution, window width and level truncation, and normalization).

[0016] c) Constructing a carotid artery segmentation model: Based on the nnU-Net segmentation network, an end-to-end 3D CTA and NCCT segmentation model was constructed. The model was trained using 100 pairs of 3D images with carotid artery region annotations to achieve automatic segmentation of the carotid artery. After training, the segmentation accuracy was evaluated on 38 independent test sets.

[0017] d) Construct a 3D CycleGAN generative model based on mask prompts and MIND structural loss: Using 1650 pairs of original NCCT and CTA images and their vessel segmentation results as training data, a mask prompt loss was designed to guide the model to focus on the core vascular region, suppressing pixel changes in non-vascular regions while encouraging pixel values ​​in vascular regions to change as needed; at the same time, a MIND structural loss for vascular regions was designed to ensure that the grayscale features of blood vessels and surrounding soft tissues in the generated NCCT images are highly consistent, thereby optimizing the anatomical details of the vascular tree branches.

[0018] e) Propose a reverse distillation strategy: Based on the bidirectional generation architecture of "NCCT→CTA" (forward) and "CTA→NCCT" (reverse), the pre-trained reverse generator (CTA→NCCT) is used as the teacher model, and the vascular silhouette and intermediate layer feature map are used as bridges. The supervision signal of reverse generation is used to assist the training of the forward generator (NCCT→CTA).

[0019] f) Training the segmentation model: Using an end-to-end training method, the segmentation model constructed in step c) is trained and optimized on 100 pairs of pre-training data.

[0020] g) Generate pseudo-labels for vessel segmentation: Apply the trained segmentation model to 1650 pairs of training set images to generate carotid artery segmentation results for each pair of NCCT and CTA images, which will serve as pseudo-labels for the subsequent generated models.

[0021] h) Two-stage training to generate the model: First, input the NCCT and CTA images in the training set and the blood vessel segmentation pseudo-labels obtained in step g) into the 3D CycleGAN generation model constructed in step d) for the first stage of training to obtain a well-trained generator A (CTA→NCCT) and an initial generator B (NCCT→CTA); then, using the back-distillation strategy of step e), freeze the generator A of the first stage, and perform the second stage back-distillation training on the generator B that inherits the parameters of the first stage, and optimize the generator B using the output silhouette of generator A and the intermediate layer feature map.

[0022] i) Testing and evaluation: The trained generative model is applied to an independent test set (38 pairs of NCCT and CTA), virtual CTA images are generated for all NCCTs in the test set, and the generation similarity index is calculated between the virtual CTA images and the corresponding real CTA images.

[0023] Specifically, in step b), of the 1788 pairs of 3D NCCT and CTA image data collected, two senior doctors from Jinling Hospital affiliated with Nanjing University School of Medicine performed pixel-level annotation of the carotid artery vessels on 138 pairs of images. The 100 pairs of data with annotations were used as pre-training data to train the carotid artery segmentation model; the remaining 38 pairs were used as an independent test set; and the 1650 pairs of data without annotations were used as the training set for the generative model. Data preprocessing specifically included: first, converting all image formats from DICOM to NIfTI; second, aligning the Origin and Direction of the NCCT images to the paired CTA images through resampling, and unifying the spatial resolution (spacing) of all images to [0.5, 0.5, 0.5]; finally, truncating the pixel values ​​of all images according to a window level of 130 and a window width of 800, and normalizing the pixel values ​​to the [0, 255] range.

[0024] Specifically, in step c), the two trained nnU-Net networks (one for NCCT and one for CTA) are both based on the nnU-Net v2 architecture and employ an encoder-decoder structure. The encoder extracts multi-scale spatial features of the carotid artery by stacking 3D convolutional layers and downsampling operations; the decoder gradually restores the feature map resolution by upsampling and skip connections, fusing shallow detail information with deep semantic information. Both use the average of Dice loss and cross-entropy (CE) loss as the loss function. The network input is a 3D image of NCCT or CTA, and the output is the corresponding vessel region segmentation result, dividing the entire image into a background region and a carotid artery region. Model training employs end-to-end supervised learning, with the loss function being a weighted combination of Dice loss and cross-entropy loss. Depth supervision is applied to multiple resolution output layers to accelerate convergence. During training, the learning rate is adaptively adjusted, and online data augmentation strategies such as random rotation, random scaling, and elastic deformation are used to improve the model's robustness to carotid artery morphological variations. After training, the constructed nnU-Net segmentation model achieved Dice coefficients of 95.56% and 91.57% for segmenting blood vessel regions in CTA and NCCT images, respectively, on independent test sets.

[0025] Specifically, in step d), the 3D CycleGAN generation model based on mask prompts and MIND structural loss uses 1650 pairs of original NCCT and CTA images and their vessel segmentation results as training data to output two generators: Generator A (CTA→NCCT) and Generator B (NCCT→CTA). First, addressing the problem that existing technologies are mostly based on 2D generation frameworks and neglect the continuity of 3D vascular space, leading to fragmentation and misalignment between slices during reconstruction, this invention uses 3D CycleGAN as the basic backbone to maintain the integrity of the three-dimensional spatial structure. Second, addressing the problem that the carotid artery accounts for less than 5% of the pixels in head and neck images, has thin and complex branches, and is easily affected by background tissue interference, causing feature learning to deviate from the core region, this invention designs and introduces a mask-guided loss, the formula of which is as follows:

[0026]

[0027]

[0028] in, and Let A and B represent the masked guided losses, respectively. and These represent pseudo-labels for vessel segmentation on CTA and NCCT images, respectively. and These represent pairs of real CTA and NCCT images, respectively. and These represent the CTA and NCCT image pairs generated by the model, respectively. This represents L1 loss. This loss simulates the imaging principles of real CTA and NCCT—contrast agents should only cause grayscale changes in the vascular region, thus guiding the model to focus on the core vascular region. While suppressing pixel changes in non-vascular regions, it encourages pixel values ​​in the vascular region to change as needed. Finally, addressing the problem in NCCT where the grayscale features of blood vessels and surrounding soft tissues are highly similar, making it difficult to preserve small vascular branches, this invention designs and introduces the vascular region MIND structural loss, the formula of which is as follows:

[0029]

[0030] in, This represents the MIND structural loss of generator A. This indicates a pseudo-label for vessel segmentation on a CTA. This represents an NCCT image generated by CTA, where MIND stands for modality-independent neighborhood descriptor. This loss drives the alignment of vascular grayscale in the NCCT image generated by CTA with the grayscale of the surrounding soft tissue, thereby optimizing the anatomical details of fine structures such as vascular dendritic branches.

[0031] Specifically, the backdistillation process in step e) uses generator A (CTA→NCCT) and generator B (NCCT→CTA) trained in step d) as initial models, and uses vascular silhouettes and intermediate layer feature maps as a bridge to perform knowledge distillation between the two generators. The specific process is as follows: First, generator A is frozen as the teacher model, and generator B, inheriting the weights from step d), is used as the student model. Training continues on the original NCCT and CTA images with 1650 pixels. Analogous to responsive distillation in knowledge distillation, and based on the fact that "CTA should only have increased grayscale in the vascular region compared to NCCT," a silhouette structure alignment loss is designed and introduced. The silhouette structure alignment is as follows: the silhouette of generator A is defined as the input CTA minus the output NCCT, and the silhouette of generator B is defined as the output CTA minus the input NCCT. By minimizing the SSIM loss between the two, the silhouette of generator B is made to approximate the silhouette of generator A, as shown in the following formula:

[0032]

[0033] in, This indicates the loss in silhouette structure alignment. and These represent real CTA and NCCT images, respectively. and These represent the CTA and NCCT images generated by the model, respectively. This represents SSIMloss. This loss prompts the silhouette of generator B (generated CTA minus input NCCT) to structurally align with the silhouette of generator A (input CTA minus generated NCCT). Secondly, analogous to feature distillation in knowledge distillation, a feature alignment loss is designed and introduced. Feature structure alignment is achieved as follows: extract the k-th layer feature map (k=1,2,3,4, K=4 layers in total) between generator A and generator B. By minimizing the SSIM loss between corresponding layer feature maps, the feature representation of generator B approximates that of generator A, as shown in the following formula:

[0034]

[0035] in, This represents the feature structure alignment loss. Represents SSIM loss, and Let K and K represent the k-th intermediate feature maps of generator A and generator B, respectively, where K=4 and k=1,2,3,4. This loss, from the intermediate feature level of the model, prompts generator B to learn from generator A, achieving a deeper level of knowledge transfer.

[0036] Beneficial Effects: This invention constructs a novel 3D angiography generation framework, achieving end-to-end generation from NCCT images to high-quality CTA images while ensuring the integrity of the overall vascular structure. This invention innovatively proposes a 3D mask-guided CycleGAN model, incorporating vascular segmentation results as auxiliary domain information into the generation process. Mask cue loss guides the model to focus on core vascular regions, while introducing MIND structural loss to precisely optimize the anatomical details of the vascular tree branches. Notably, this invention is the first to propose the concept of backdistillation, using vascular silhouettes and intermediate layer feature maps as a bridge to achieve knowledge transfer between generators with completely opposite inputs and outputs, further improving the image quality and anatomical fidelity of virtual CTA. Ultimately, during inference, only NCCT images need to be input to automatically generate high-quality virtual CTA images that meet clinical diagnostic needs, balancing generation performance and clinical practicality, providing a new technical path for contrast-free vascular imaging. Attached Figure Description

[0037] Figure 1 This is an overall framework diagram of the method proposed in this invention.

[0038] Figure 2 This is a comparison image of the blood vessel segmentation results and doctor annotations within the framework of the method proposed in this invention. The first column shows the original NCCT and CTA images, the second column shows the model segmentation results, and the third column shows the doctor's pixel-level annotations.

[0039] Figure 3 This is the result image generated in the first stage of the method framework proposed in this invention. The first column is the original NCCT, the second column is the NCCT generated by generator A, the third column is the CTA generated by generator B, and the fourth column is the original CTA.

[0040] Figure 4 This is a CTA generation result diagram of the second stage in the method framework proposed in this invention. The first column is the original NCCT, the second column is the CTA generated by generator B, and the third column is the original CTA.

[0041] Figure 5 The graphs are the silhouettes, threshold graphs, and maximum connected component graphs of the generated image and the original image in this invention. The first column is the silhouette of the generated image and the original image, the second column is the threshold result of the silhouette image, the third column is the maximum connected component result of the threshold result, and the fourth column is the pixel-level annotation of the doctor's blood vessels.

[0042] Figure 6 This is a visual comparison of the generation results of the method of the present invention and other generation methods on an independent test set. Detailed Implementation

[0043] The objectives, technical solutions, and advantages of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] This invention addresses four core technical bottlenecks in existing generative models: limitations in generation paradigms, extremely low target proportion, data alignment bias, and difficulty in boundary differentiation. It proposes a mask-guided and reverse-distillation-based method for enhancing three-dimensional angiography of the carotid artery using non-contrast CT. This method requires only input of NCCT images to generate high-quality CTA images that meet clinical needs. The overall flowchart of this invention is shown below. Figure 1 As shown, the specific steps include:

[0045] a) Data collection: 1788 paired CTA and NCCT image data were collected from Jinling Hospital affiliated with Nanjing University School of Medicine. First, doctors selected and labeled the carotid artery region in 138 pairs of images. After labeling, 100 pairs were used as the training set for the deep network segmentation model, and the remaining 38 pairs were used as the independent test set; the remaining 1650 pairs of unlabeled data were used as the training set for the deep network generation model.

[0046] b) Establishing a CTA and NCCT carotid artery segmentation model: Based on the nnU-Net v2 segmentation network, an end-to-end 3D CTA and NCCT image segmentation model was constructed. The model training hyperparameters are shown in Table 1. After training, carotid artery segmentation was performed on the generated model training set and independent test set for CTA and NCCT images, respectively. The segmentation results are shown in Table 1. Figure 2 As shown.

[0047] Table 1. Hyperparameters for segmentation model training

[0048]

[0049] (c) Establishing a 3D Cyclegan generative model based on mask prompts and MIND structural loss: Using 1650 pairs of original NCCT and CTA images and their vessel segmentation pseudo-labels as training data, the following mask prompt loss and vessel region MIND structural loss were designed. The training process for this stage is as follows: Figure 1 The subgraph is shown in Step 1.

[0050]

[0051]

[0052]

[0053] The model training hyperparameters are shown in Table 2. Through this training stage, a well-trained generator A (CTA→NCCT) and an initially coarse generator B (NCCT→CTA) are obtained. The generation effect in this stage is as follows: Figure 3 As shown.

[0054] Table 2. Stage 1 Generative Model Training Hyperparameters

[0055]

[0056] (d) Establishing the backdistillation model: Using generator A and generator B obtained in stage one as the initial model, generator A is fixed, and the following blood vessel silhouette loss and intermediate layer feature map loss are designed to assist generator B in continuing training. The training process in this stage is as follows: Figure 1 Subgraph Step2 is shown.

[0057]

[0058]

[0059] The model training hyperparameters are shown in Table 3. The final optimized generator B's generation effect is shown in Table 3. Figure 4 As shown.

[0060] Table 3. Stage 2 Generative Model Training Hyperparameters

[0061]

[0062] (e) Inference and Quantitative Evaluation: After training, the generator B, optimized in step d), is applied to an independent test set (38 pairs of NCCT and CTA images). For each NCCT image in the test set, the generator B is used to generate a corresponding virtual CTA image. To quantitatively evaluate the generation effect, the generated virtual CTA is compared with the real CTA image, and evaluation metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are calculated. Simultaneously, silhouette analysis, threshold segmentation, and maximum connected component extraction are performed on the generated results, and the results are compared and verified with the vascular regions annotated by the physician. Figure 5 As shown.

[0063] To more clearly illustrate the feasibility and superiority of this invention, Table 4 presents the quantitative metrics of the proposed method (Step 1 and Step 2) on an independent test set, and compares them numerically with various 2D and 3D generation methods (including CycleGAN, pix2pix, CTA-GAN, DDIM, Fast-DDPM, AEGIS, etc.). Evaluation metrics include PSNR, SSIM, multi-scale structural similarity (MS-SSIM), and vascular region mutual information (MI). As can be seen from the table, the Step 2 method proposed in this invention outperforms the comparative methods in all metrics, especially showing significant improvements in SSIM and MS-SSIM metrics, which are closely related to anatomical structures, thus verifying the effect of the backdistillation strategy on improving the quality of the generated images.

[0064] Table 4 Quantitative Indicators of Generative Model Results

[0065]

[0066] Furthermore, to visually demonstrate the improved effect of the method of the present invention, Figure 6 The virtual CTA images generated by different methods (including 2D / 3D CycleGAN, Step1, and Step2) under the same NCCT input were compared with real CTA images. As can be seen from the figures, the blood vessel structure generated by the Step2 method of this invention is more continuous, the edges are clearer, and the small branches are more completely preserved, which is consistent with the trend of the quantitative indicators in Table 4.

[0067] This invention first processes CTA and NCCT image pairs and combines them with pixel-level annotations of carotid arteries by doctors to construct a 3D segmentation pre-training dataset. Based on this dataset, automatic carotid artery segmentation models for CTA and NCCT images are trained using the nnU-Net deep network. On independent test sets, the two segmentation models achieve Dice coefficients of 95.56% and 91.57% for carotid artery segmentation in CTA and NCCT images, respectively. Subsequently, pseudo-labels for vessel segmentation are generated for 1650 training image pairs using the trained segmentation models. Mask cueing loss and MIND structural loss are designed and integrated into the 3D CycleGAN generation model for the first stage of training, obtaining a preliminary forward generator B (NCCT→CTA) and a high-performance reverse generator A (CTA→NCCT). Finally, based on the results of the first stage of training, generator A is fixed, and a backdistillation concept is proposed. By introducing vessel silhouette loss and intermediate layer feature map loss, generator B is further optimized using generator A to complete the second stage of training. In summary, through the above steps, this invention achieves fully automated, high-quality generation from NCCT to virtual CTA, obtaining three-dimensional vascular images that meet clinical diagnostic needs without the need for contrast agents, thus verifying the effectiveness and superiority of the proposed method.

Claims

1. A method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation, characterized in that, Includes the following steps: a) Data collection and preprocessing: Collect paired CTA and NCCT images; divide the collected data into pre-training data, training set, and test set according to whether they contain vascular annotations, and perform preprocessing. b) Constructing a carotid artery segmentation model: Based on the nnU-Net segmentation network, an end-to-end 3D CTA and NCCT image segmentation model is established for accurate segmentation of the carotid artery. c) Construct a 3D CycleGAN generative model based on mask cueing and MIND structural loss: Based on NCCT, CTA images and their vessel segmentation results, mask cueing loss and MIND structural loss for vascular regions are designed to guide the generative model to focus on the core vascular region and optimize the anatomical details of the vascular tree branches. d) Propose a reverse distillation strategy: Based on the bidirectional generation architecture of "NCCT→CTA" and "CTA→NCCT", using vascular silhouettes and intermediate layer feature maps as bridges, and using the supervision signal generated by reverse generation to assist the training of the forward model; e) Training the segmentation model: Using an end-to-end training method, the segmentation model constructed in step b) is trained and optimized on the pre-training data; f) Generate pseudo-labels for blood vessel segmentation: After the segmentation model is trained, it is applied to the training set to generate pseudo-labels for blood vessel segmentation for all paired CTA and NCCT images in the training set. g) Training the generative model: Input the CTA and NCCT images in the training set and the blood vessel segmentation pseudo-labels obtained in step f) into the 3D CycleGAN generative model constructed in step c) for the first stage of training to obtain a well-trained generator A and an initial generator B; then, using the back distillation strategy in step d), freeze generator A and use generator A to assist generator B in the second stage of training. h) Testing and evaluation: After training, the generative model obtained in step g) is applied to an independent test set to generate virtual CTA images for all NCCT images in the test set, and the generation similarity index is calculated between the virtual CTA images and the corresponding real CTA images.

2. The method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation according to claim 1, characterized in that: In step a), the collected NCCT and CTA paired data are divided into the following categories according to whether they contain vascular annotations: 100 pairs of pre-training data with vascular annotations, 1650 pairs of training data without vascular annotations, and 38 pairs of test data with vascular annotations. The preprocessing specifically includes: first, converting all image pairs from DICOM format to NIfTI format, then resampling to a unified Origin, Direction, and spatial resolution, where the spatial resolution Spacing = [0.5, 0.5, 0.5]. Finally, truncating the pixel values ​​of all images according to a window level of 130 and a window width of 800, and normalizing them to the interval [0, 255].

3. The method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation according to claim 1, characterized in that: In step b), the carotid artery segmentation model is constructed based on the nnU-Net v2 network and adopts an encoder-decoder structure; the encoder extracts the multi-scale spatial features of the carotid artery vessels step by step through stacked three-dimensional convolutional layers and downsampling operations. The decoder gradually recovers the feature map resolution through upsampling and skip connections, fusing shallow detail information with deep semantic information; the model training adopts end-to-end supervised learning, the loss function is a weighted combination of Dice loss and cross-entropy loss, and deep supervision is applied to multiple resolution output layers; During training, the learning rate is adaptively adjusted, and online data augmentation strategies such as random rotation, random scaling, and elastic deformation are used to improve the model's robustness to variations in carotid artery morphology.

4. The method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation according to claim 1, characterized in that: In step c), the 3D CycleGAN generative model based on mask prompts and MIND structural loss uses 3D CycleGAN as its basic framework and designs mask guidance loss and MIND structural loss, with the specific formulas as follows: in, and Let A and B represent the masked guided losses, respectively. and These represent pseudo-labels for vessel segmentation on CTA and NCCT images, respectively. and These represent pairs of real CTA and NCCT images, respectively. and These represent the CTA and NCCT image pairs generated by the model, respectively. Indicates L1 loss, Let MIND represent the structural loss of generator A, where MIND stands for modality-independent neighborhood descriptor.

5. The method for enhancing carotid artery non-contrast CT three-dimensional angiography based on mask guidance and reverse distillation according to claim 1, characterized in that: In step d), the reverse distillation concept uses vascular silhouettes and intermediate layer feature maps as a bridge between generators A (CTA→NCCT) and B (NCCT→CTA), which operate in opposite directions. During training, generator A is frozen as the teacher model, and generator B is used as the student model. The silhouette structure alignment is achieved as follows: the silhouette of generator A is defined as the input CTA minus the output NCCT, and the silhouette of generator B is defined as the output CTA minus the input NCCT. By minimizing the SSIM loss between the two, the silhouette of generator B is made to approximate the silhouette of generator A, as shown in the following formula: in, This indicates the loss in silhouette structure alignment. Represents SSIM loss, and These represent pairs of real CTA and NCCT images, respectively. and These represent the CTA and NCCT image pairs generated by the model, respectively. The feature structure alignment is achieved as follows: The k-th layer feature map between generator A and generator B is extracted, where k = 1, 2, 3, 4, for a total of K = 4 layers; by minimizing the SSIM loss between corresponding layer feature maps, the feature representation of generator B approximates that of generator A, as shown in the following formula: in, This represents the feature structure alignment loss. Indicates SSIM loss, and These represent the intermediate feature maps of the k-th layer of generator A and generator B, respectively.

Images