Coronary artery stenosis dynamic quantization identification method based on diffusion model and YOLO hypergraph optimization

Abstract

The invention discloses a dynamic quantization recognition method for coronary stenosis based on a diffusion model and YOLO hypergraph optimization, and provides a self-supervised blood vessel segmentation method for solving the problems that manual annotation of a blood vessel mask is challenging and needs to consume a large number of resources due to the fact that blood vessel branches are fine and the structure is complex. The method comprises the following steps of: performing blood vessel segmentation by adopting a diffusion adversarial representation learning (DARL) model based on Transform; and aiming at the limitation of a traditional YOLO model in the aspect of feature fusion, difficulty in fully capturing a cross-level and cross-position complex feature relationship and incapability of providing accurate information of a narrow position, a Hyper-YOLO algorithm for locally enhancing attention is provided, so that blood vessel stenosis lesion identification can be effectively realized, and a good identification effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and particularly relates to a method and system for dynamically quantifying and identifying coronary artery stenosis based on diffusion model and YOLO hypergraph optimization. Background Technique

[0002] Coronary heart disease (CHD) is one of the main causes of death globally. Its main characteristics are the hardening and stenosis of the coronary arteries, leading to myocardial ischemia, and then triggering a series of clinical symptoms such as chest pain, heart failure, etc., and even endangering life in severe cases. With the increasing incidence of cardiovascular diseases year by year, the early diagnosis and precise treatment of coronary heart disease have become important topics in medical research and clinical practice. Imaging examinations, especially angiography (such as coronary angiography) and CT angiography (CTA), provide important bases for the diagnosis and treatment decision-making of coronary heart disease. Angiography images can intuitively reflect the morphology, width, and blood flow status of blood vessels, so they are widely used in the diagnosis and interventional treatment of coronary heart disease. However, the analysis of angiography images is challenging, especially in the precise identification and quantitative analysis of vascular stenosis lesions, which are often affected by factors such as noise, low contrast, and complex vascular branches.

[0003] Vascular segmentation in angiography images is the basis of image analysis and determines the accuracy of subsequent lesion detection and quantitative analysis. Due to the complexity of the vascular structure, the diversity of branches, and the weakness of the vascular wall, the segmentation task often faces problems such as fuzzy boundaries, background interference, and noise effects. Traditional image processing methods are difficult to meet the requirements of high-precision segmentation. Common vascular segmentation methods such as the Frangi algorithm based on the Hessian matrix, the method based on principal component analysis (PCA), and the deep learning method based on U-Net, although achieving good results in some scenarios, their performance significantly declines when facing small blood vessels, complex vascular branches, or low-quality angiography images. Especially in complex lesions and branch structures, the automatic annotation and segmentation of blood vessels still rely on a large amount of manual participation, increasing the burden of clinical work and the possibility of errors. Therefore, developing an automated method that can process complex vascular images and improve the accuracy of vascular segmentation has become an urgent need for vascular image analysis.

[0004] Meanwhile, in the diagnosis and treatment of coronary heart disease, accurately identifying vascular stenosis lesions is of crucial importance. The degree of vascular stenosis directly affects the choice of treatment plan and the prognosis of patients. The fractional flow reserve (FFR) currently used in clinical practice is an important method for evaluating the degree of vascular stenosis. FFR evaluates the impact of vascular stenosis on blood flow by measuring the hemodynamic characteristics of the coronary artery. However, this method still has its limitations. FFR can only provide the physiological impact value of stenosis, but cannot accurately locate the position of the stenosis area, and depends on the parameters and conditions during clinical operations, often resulting in a certain impact on the accuracy of diagnosis. In addition, FFR detection requires invasive operations, increasing the risk to patients and medical costs.

[0005] Therefore, in recent years, automated analysis methods based on medical images have received extensive attention, especially the application of deep learning in medical image processing. Models based on convolutional neural networks (CNNs) have been widely used in the segmentation of vascular images and lesion recognition, achieving certain results. However, traditional deep learning methods still face problems such as complex background noise, thin vascular branches, and blurred lesion areas in vascular images, resulting in limited accuracy of segmentation results. In response to these problems, many improvement strategies have been proposed in recent years, such as introducing attention mechanisms to enhance attention to key regions, or using generative adversarial networks (GANs) to improve the expression ability of image features, thereby optimizing the effects of vascular segmentation and lesion detection.

[0006] In deep learning methods, image processing methods based on Transformer have also received extensive attention in recent years. Through the self-attention mechanism, Transformer can effectively capture long-range dependencies in images, thereby improving the accuracy of image segmentation and object detection. In vascular segmentation and lesion recognition, the Transformer model can better handle the complexity of vascular morphology and the diversity of lesion areas, avoiding the problem of information loss that traditional convolutional neural networks may encounter when dealing with complex structures.

[0007] The accurate identification of vascular stenosis lesions is also a research hotspot in the field of deep learning. Existing lesion recognition technologies mainly focus on the auxiliary calculation of fractional flow reserve (FFR) and the automatic annotation of vascular lesion areas. Although object detection models based on YOLO, Faster R-CNN, etc. can identify vascular stenosis lesions, the capture of details is still limited, and in complex coronary artery branches and vascular overlap areas, the effects of traditional object detection methods are often unsatisfactory. Therefore, combining attention mechanisms with efficient detection frameworks and optimizing existing algorithms to improve the recognition accuracy of vascular stenosis lesions has become the focus of current technical research.

[0008] Therefore, in view of the deficiencies of existing vascular segmentation and lesion detection technologies, there is an urgent need in the art for a technical solution that can solve problems such as low accuracy of vascular segmentation and difficulty in accurately identifying lesion areas. By combining advanced image processing and deep learning technologies to improve the accuracy of segmentation and lesion recognition, it can effectively promote the intelligent diagnosis and treatment process of coronary heart disease, reduce the dependence on manual annotation, optimize clinical decision-making, and ultimately provide more accurate and personalized treatment plans for patients. Summary of the Invention

[0009] The object of the present invention is to provide a method and system for dynamically quantifying and identifying coronary artery stenosis based on a diffusion model and YOLO hypergraph optimization, which solves problems such as complex backgrounds, noise, and low contrast in angiography images, thereby effectively improving the accuracy of vascular segmentation and enhancing the detection ability of lesion areas.

[0010] To solve the above problems, the technical solution of the present invention is as follows: A method for dynamically quantifying and identifying coronary artery stenosis based on a diffusion model and YOLO hypergraph optimization, comprising: Preprocess the angiography image containing noise, learn the denoising process through a diffusion module, gradually recover from the noisy image to the original image, and extract the background signal; Adopt a diffusion adversarial representation learning model based on Transformer, for the background signal, generate a vascular segmentation mask through a generation module and adversarial training to achieve accurate segmentation of blood vessels; Adopt the Hyper-YOLO algorithm with local enhanced attention, insert an attention module after the convolutional layer to enhance the feature response to the vascular stenosis area, and combine hypergraph calculation to model the high-order associations in the feature map to achieve accurate identification of vascular stenosis lesions; Visualize the recognition results, including the annotation of the stenosis location and severity, to provide a quantitative basis for clinical diagnosis.

[0011] According to an embodiment of the present invention, the preprocessing of the angiography image containing noise includes: Based on the denoising diffusion probability model, gradually reduce the noise through multiple time steps, starting from the image with the highest noise level, and gradually estimate the denoised image at each intermediate step until the original image is restored; During the denoising process, gradually identify and restore the background signal in the image to enhance the ability to recognize the vascular morphology.

[0012] According to an embodiment of the present invention, the diffusion adversarial representation learning model based on Transformer includes: Diffusion module: used to estimate latent features, gradually remove noise through the reverse diffusion process, and learn the background signal of the image; Generation module: Introduce the Transformer structure to enhance global feature learning, capture the long-range dependencies of blood vessels, and generate blood vessel segmentation masks and synthetic angiograms; Adversarial training unit: Through the adversarial learning of the generator and discriminator, generate realistic blood vessel images and segmentation masks, and improve the model's ability to capture the semantic information of blood vessels.

[0013] According to an embodiment of the present invention, a Transformer encoder module is introduced into the encoder of the generation module, and the preliminarily extracted features of the input image are sent into the Transformer encoder module for global context modeling; A Transformer decoder module is introduced into the decoder, and the features output by the encoder are processed by the Transformer decoder, and the image is gradually restored through upsampling and convolutional layers.

[0014] According to an embodiment of the present invention, the generation module is composed of multiple residual blocks with SPADE layers, and the SPADE layer normalizes the feature map in different ways according to whether the input data includes a segmentation mask; The generation module generates a blood vessel segmentation mask without using the SPADE layer according to the input noisy angiogram and latent features; and / or synthesizes an angiogram using the SPADE layer according to the input segmentation mask, the noisy background and its latent features.

[0015] According to an embodiment of the present invention, the Transformer-based diffusion adversarial representation learning model constructs a generator for receiving the noisy latent representations at different stages in the diffusion process as inputs and outputting the generated latent representations; the generated latent representations are close to the original latent representations; Construct a discriminator for distinguishing the latent representations generated by the generator and the real original latent representations; Through the way of adversarial training, alternately update the parameters of the generator and the discriminator to make them reach a balance.

[0016] According to an embodiment of the present invention, the local enhanced attention Hyper-YOLO algorithm includes: Insert a local attention module after the convolutional layer of Hyper-YOLO to enhance the feature response to the key regions of blood vessels and improve the localization sensitivity to narrow regions; Based on the hypergraph convolution HyperC2Net, model the complex high-order associations in the feature map, realize the information propagation across levels and positions, and improve the lesion recognition accuracy; Combine the blood vessel segmentation mask information, and detect narrow lesion regions of different sizes through a multi-scale prediction head to achieve multi-object detection.

[0017] According to an embodiment of the present invention, further including inserting a local attention module after the convolutional layer of Hyper-YOLO: Inserting a local attention module after the key convolutional layers of the backbone network and the feature pyramid of Hyper-YOLO; The local attention module receives the feature map containing the key regions of blood vessels output by the convolutional layer, divides the feature map into k×k local windows, and independently calculates the attention weights for each window; Multiplying the attention weights and the feature map element by element to obtain an enhanced feature map.

[0018] According to an embodiment of the present invention, further including detecting narrow lesion regions of different sizes through a multi-scale prediction head: Inputting the enhanced feature map into multiple prediction heads of Hyper-YOLO, and outputting the bounding boxes, confidence levels, and classification results of each prediction head; Calculating the intersection over union (IoU) for all bounding boxes. When the IoU of two bounding boxes is higher than the threshold, retain the bounding box with the higher confidence level; after pairwise comparison, remove redundant boxes and retain the bounding box with the highest confidence level as the output result; Mapping the output bounding boxes from the scaled feature map back to the original image space.

[0019] A coronary stenosis dynamic quantification and recognition system based on a diffusion model and YOLO hypergraph optimization, including: An image preprocessing module: used to preprocess the angiography image containing noise, learn the denoising process through the diffusion module, gradually recover from the noisy image to the original image, and extract the background signal; A blood vessel segmentation module: adopting a diffusion adversarial representation learning model based on Transformer, for the background signal, generating a blood vessel segmentation mask through the generation module and adversarial training to achieve precise segmentation of the blood vessels; A lesion recognition module: adopting a Hyper-YOLO algorithm with local enhanced attention, by inserting an attention module after the convolutional layer, enhancing the feature response to the blood vessel stenosis region, and combining hypergraph calculation HyperC2Net to model the high-order associations in the feature map to achieve precise recognition of blood vessel stenosis lesions; A result output module: visually displaying the recognition results, including the annotation of the stenosis location and severity, providing a quantitative basis for clinical diagnosis.

[0020] Due to the adoption of the above technical solutions, the present invention has the following advantages and positive effects compared with the prior art: In an embodiment of the present invention, a coronary artery stenosis dynamic quantification and recognition method based on a diffusion model and YOLO hypergraph optimization aims at the problem that it is challenging and resource-consuming to manually annotate vascular masks due to the fine and complex structure of blood vessel branches. A self-supervised vascular segmentation method is proposed, that is, a diffusion adversarial representation learning (DARL) model based on Transformer is used for vascular segmentation; and aiming at the limitation of the traditional YOLO model in feature fusion, which is difficult to fully capture complex feature relationships across levels and positions, and cannot provide accurate information on the stenosis location, a Hyper-YOLO algorithm with local enhanced attention is proposed, which can effectively realize the recognition of vascular stenosis lesions and achieve good recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of a coronary artery stenosis dynamic quantification and recognition method based on a diffusion model and YOLO hypergraph optimization in an embodiment of the present invention; Figure 2 is a schematic structural diagram of the DARL model in an embodiment of the present invention; Figure 3 is a schematic structural diagram of the MANet model in an embodiment of the present invention; Figure 4 is a system block diagram of a coronary artery stenosis dynamic quantification and recognition system based on a diffusion model and YOLO hypergraph optimization in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The following further describes in detail a coronary artery stenosis dynamic quantification and recognition method and system based on a diffusion model and YOLO hypergraph optimization proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will be clearer according to the following description and the claims.

[0023] DARL, namely diffusion adversarial representation learning, is an adversarial vascular segmentation method based on diffusion. It includes a diffusion module and a generation module. This method uses the diffusion module to learn background signals, thereby effectively improving the performance of vascular segmentation and noise resistance, and combines the synthesis of a given contrast foreground image and a semantic image of a fractal vascular mask to estimate the vascular segmentation mask. The vascular fractal mask is provided to the generation module, and image segmentation is performed according to the semantic layout. The DARL model can effectively learn the semantic information of blood vessels without a large amount of labeled data and achieve accurate segmentation of blood vessels in medical images.

[0024] The YOLO (You Only Look Once) series is a mainstream method in the field of object detection, known for its efficiency and real-time performance. Hyper-YOLO is a kind of object detection method based on hypergraph computing in the YOLO series. It integrates hypergraph computing into the object detection network for the first time, models the complex high-order associations in the feature map, realizes the cross-level and cross-position propagation of high-order information, and has a more significant improvement in the recognition of small and medium-sized objects.

[0025] In this embodiment, aiming at the problem that it is challenging and resource-consuming to manually annotate vascular masks due to the fine and complex structure of blood vessel branches, a self-supervised vascular segmentation method is proposed, that is, using a diffusion adversarial representation learning (DARL) model based on Transformer for vascular segmentation; and aiming at the limitations of traditional YOLO models in feature fusion, which are difficult to fully capture the complex feature relationships across levels and positions, and cannot provide accurate information on narrow positions, a Hyper-YOLO algorithm with local enhanced attention is proposed, which can effectively identify vascular stenosis lesions and achieve good recognition results.

[0026] Please refer to Figure 1 , the coronary artery stenosis dynamic quantification and recognition method based on the diffusion model and YOLO hypergraph optimization provided in this embodiment includes the following steps: S1: Preprocess the noisy angiography image, learn the denoising process through the diffusion module, gradually recover from the noisy image to the original image, and extract the background signal; S2: Use a diffusion adversarial representation learning model based on Transformer to generate a vascular segmentation mask for the background signal through the generation module and adversarial training, realizing the accurate segmentation of blood vessels; S3: Use the Hyper-YOLO algorithm with local enhanced attention to enhance the feature response to the vascular stenosis area by inserting an attention module after the convolutional layer, and combine hypergraph computing to model the high-order associations in the feature map, realizing the accurate recognition of vascular stenosis lesions; S4: Visualize the recognition results, including the annotation of the stenosis position and severity, providing a quantitative basis for clinical diagnosis.

[0027] In step S1, in the noisy angiography image, the denoising process is learned through the diffusion module, and the image is gradually recovered from the noisy image to the original image. This process enables the model to effectively learn the background signal of the image. The denoising process is as follows: In the diffusion module, the denoising diffusion probabilistic diffusion model (DDPM) is used for learning. The diffusion model is a generative model based on the physical diffusion process. It restores a noisy image to a realistic image through a step-by-step process. This process can be viewed as a reverse diffusion process: starting from a noisy image, the noise is gradually removed and the original image is restored. The model's learning process involves learning how to gradually remove noise through backpropagation.

[0028] In practice, the diffusion module gradually reduces noise over multiple time steps. ( is the maximum time step), the network gradually estimates the denoised image at each intermediate step , until it is finally restored to a close to original image .

[0029] Specifically, at each time step , the network is based on the current angiography noise image Prediction noise (Subject to standard normal distribution ), and then estimate the denoised image by the following formula : (1) in, is a time step A related parameter that controls the degree of noise.

[0030] As time steps progress, the noise gradually decreases and the image becomes closer and closer to the original noise-free image. This gradual denoising process enables the model to effectively capture various features in the image, including background signals.

[0031] During the denoising process, the diffusion model not only removes noise but also gradually identifies the background signal of the image through learning at each time step. Specifically, through the reverse diffusion process, the model can simultaneously remove noise and identify key information in the image, such as the shape and structure of blood vessels. This includes: 1. Restoring background signals: Background signals in an image are relatively stable and do not change significantly. They are typically the main structure or the background of the image. Through the denoising process, the model can recover these stable signals from noisy images, helping the model to recognize the true shape of blood vessels without being disturbed by noise.

[0032] 2. Background Learning: During training, the model gradually learns how to identify background signals in an image and separate them from noise. Because background signals in images are typically stable, the diffusion model can gradually strengthen its recognition of background information through multiple iterations, ultimately making the background signal clearer and reducing the impact of noise.

[0033] Denoising noisy angiography images using a diffusion model effectively restores the original image structure and enhances the clarity of background signals. The diffusion process, through a combination of forward and backward diffusion, gradually removes noise and restores the original image. Ultimately, the model effectively learns the background signal in the image, thereby improving the accuracy of vascular lesion detection.

[0034] In step S2, a Transformer-based diffusion adversarial representation learning model is used to generate a blood vessel segmentation mask for the background signal through a generation module and adversarial training, thereby achieving accurate blood vessel segmentation. The Transformer-based diffusion adversarial representation learning model includes: Diffusion module: used to estimate latent features, gradually remove noise through the reverse diffusion process, and learn the background signal of the image; Generation module: Introducing the Transformer structure to enhance global feature learning, capture the long-range dependencies of blood vessels, generate vessel segmentation masks and synthesize angiography images; Adversarial training unit: Through adversarial learning between the generator and the discriminator, realistic vascular images and segmentation masks are generated, improving the model's ability to capture vascular semantic information.

[0035] Specifically, the functions of the diffusion module are as described in step S1 and will not be repeated here. In the generation module, the Transformer replaces the traditional convolutional layer to capture global context information. Specifically, a Transformer encoder module can be introduced in the encoder part of the generation module, and a Transformer decoder module can be introduced in the decoder part. This design can effectively model the global characteristics of the image, improving denoising and restoration quality.

[0036] In the encoder, the input image is passed through the convolutional layer to extract preliminary features, and the extracted features are input into the Transformer encoder for global context modeling.

[0037] Self-Attention Mechanism: (2) Q is the query, K is the key, V is the matrix, The dimension of the key.

[0038] Multi-head attention mechanism: (3) (4) h is the number of attention heads, is the output weight matrix.

[0039] In the decoder, the features output by the encoder are processed through the Transformer decoder, and the image details are gradually restored through upsampling and convolutional layers.

[0040] Self-attention mechanism: (5) Cross-attention mechanism (6) X is the encoder input feature matrix, and Y is the decoder input feature matrix.

[0041] By introducing the Transformer structure into the generation module, the denoising quality can be effectively improved, the image restoration effect can be enhanced, and the long-range dependencies of structures such as blood vessels can be captured. This method improves the segmentation effect by enhancing global feature learning and paying attention to the global blood vessel morphology.

[0042] In this embodiment, the images before and after injecting the contrast agent into the blood vessels are respectively referred to as the background and the angiogram. The DARL model is trained on unpaired angiogram images and background images. Specifically, the model consists of: a diffusion module for estimating latent features; a generation module for estimating the blood vessel segmentation mask and synthesizing the angiogram; and two discriminators for distinguishing between real and fake images of the blood vessel mask and the angiogram. When generating the angiogram image, the vascular fractal mask is provided to the generation module to perform image synthesis according to the semantic layout.

[0043] Performing adversarial training using the DARL model includes the following steps: Construct a generator network whose goal is to learn the latent representation of recovering the original data from noise. The generator receives the noisy latent representations at different stages of the diffusion process as inputs and tries to generate outputs as close as possible to the original latent representation.

[0044] At the same time, construct a discriminator network for distinguishing the latent representation generated by the generator and the real original latent representation. The discriminator receives the latent representation as input and outputs a probability value indicating the probability that the input latent representation is real data.

[0045] By means of adversarial training, the parameters of the generator and discriminator are updated alternately. During the training process, their parameters are adjusted according to the loss functions of the generator and discriminator, so that the capabilities of both are continuously improved and finally reach a balanced state.

[0046] In the generation module, self-supervised vascular segmentation is adopted. For the vascular segmentation task, due to the complex vascular structure with many tiny branches, it is difficult to obtain fine-grained labels for supervised learning. In fact, unsupervised learning methods trained with natural images have strong generalization ability on unseen datasets, so they can easily adapt to the medical image segmentation task. As an unsupervised learning, self-supervised learning is introduced to effectively learn the target representation in various medical image segmentation tasks by using the supervision labels self-generated by the data itself, and its potential has been demonstrated.

[0047] A generation module with switchable SPADE layers is adopted, as Figure 2 shown. The proposed generation module consists of N residual blocks with switchable SPADE layers. The (A) and (B) paths of the model are implemented simultaneously by sharing learnable parameters except for the S-SPADE layer. Then, let be the feature map in the ResnetBlock, where M, C, H, and W are the sizes of the batch, channels, height, and width respectively. The switchable SPADE layer normalizes the feature map in different ways according to the presence of the input mask s: (7) where, is instance normalization. Therefore, when our model is given a fake vascular mask, the calculation of SPADE is as follows: (8) where, represents the th element of the feature tensor , ([[]] , ) are the mean and standard deviation of the feature map in channel c, is the modulation parameter learned during training.

[0048] Therefore, in the (A) path, given the noisy angiogram and the latent feature , the generation module G estimates the vascular segmentation mask without using SPADE: , (9) On the other hand, when the segmentation mask In the (B) path, the generation module uses the noise background and its potential features to synthesize a fake angiogram : (10) The DARL model does not require an iterative inverse process and is similar to recent diffusion-based unsupervised learning methods. Specifically, once the proposed DARL is trained, in inference, the vascular segmentation mask of the angiogram from the (A) path can be obtained in one step. For the noisy angiogram given by the forward diffusion process, the model of this embodiment uses the potential features estimated from the diffusion module to provide a vascular segmentation mask. The model of this embodiment can generate a segmentation mask for any noise level within a certain range (i.e., ). However, since angiogram images can be regarded as one of the clean target images, the closer to zero, the better the vascular segmentation performance. Therefore, by setting to test the model of this embodiment. In step S3, the Hyper-YOLO algorithm with local enhanced attention is adopted. By inserting an attention module after the convolutional layer, the feature response to the vascular stenosis area is enhanced, and the high-order correlation in the feature map is modeled by combining hypergraph calculation to achieve accurate identification of vascular stenosis lesions.

[0049] Hyper-YOLO is an optimization of the YOLO algorithm, aiming to improve the accuracy and speed of object detection. YOLO divides the image into grids, and each grid is responsible for predicting the bounding box and category and performs parallel processing. Hyper-YOLO has been optimized in feature extraction and usually uses a deeper or wider network structure for more effective feature learning.

[0050] Insert a local attention module after the key convolutional layers of the backbone network (such as DarkNet) and the feature pyramid (FPN) of Hyper-YOLO to enhance the focusing ability on the vascular stenosis area. The specific steps are as follows:

[0051] Obtain the feature map output by the convolutional layer ; Divide the feature map into local windows, and calculate the attention weights independently for each window. Among them, the attention weight generation:

[0052] (1) Local feature aggregation: Perform mean pooling on the features within each window to generate a local context vector ; ; (2) Weight calculation: Generate attention weights through a fully connected layer (or lightweight convolution): (11) Among them, Multiply the attention weights element-wise with the original feature map to obtain the enhanced feature map.

[0053] Note: Module insertion position: (1) Backbone network: Insert the local attention module after the residual block of DarkNet (such as after the C3 layer); (2) Feature Pyramid Network (FPN): Add the module after cross-scale feature fusion to enhance the detection of narrow regions at multiple scales.

[0054] Lightweight design (1) Grouped convolution: Use grouped convolution in and to reduce the number of parameters; (2) Window overlap: Adopt a sliding window (stride s < k) to increase the continuity of local context.

[0055] The local attention mechanism can focus on the key regions in the image, thereby focusing on the local features of lesions such as vascular stenosis. Since vascular lesions are usually more obvious in local regions, the local attention mechanism helps to extract the detailed information of these key regions and improve the detection accuracy of the lesion regions; in the identification of vascular stenosis lesions, there are strong local feature differences in the vascular morphology changes and lesion locations. The local attention mechanism can help the model better adapt to the changes of different lesion types and different vascular morphologies by adaptively weighting the feature maps at different positions, thereby improving the robustness of the model in diverse cases; vascular stenosis lesions usually present as local structural changes, while the background in the image often contains a lot of information unrelated to the lesions. The local attention mechanism can effectively suppress these background interferences, focus on the vascular lesion regions, and reduce the negative impact of non-target regions on the detection results; in medical images, especially in the detection task of vascular lesions, accuracy and sensitivity are crucial. The local attention mechanism can enhance the sensitivity of the model to small and difficult-to-identify lesion regions, thereby improving the detection accuracy of vascular stenosis lesions; Hyper-YOLO itself already has strong multi-scale feature extraction capabilities. After combining with the local attention mechanism, it can dynamically adjust the attention distribution at each scale, making the model have good adaptability to different types of vascular lesions. At the same time, the efficiency of the YOLO architecture itself can ensure real-time detection on large-scale medical image datasets.

[0056] YOLO series methods usually consist of two main components: the backbone and the neck. The backbone is responsible for extracting basic visual features, while the neck helps fuse multi-scale features for final object detection. Represent the three-scale outputs of the neck as {N3, N4, N5}, corresponding to small-scale, medium-scale, and large-scale detections respectively. In the feature extraction stage of the backbone network, it is further divided into five stages: {B1, B2, B3, B4, B5}, representing features at different semantic levels respectively. The larger the number, the more advanced semantic features are extracted by deeper layers of the network.

[0057] Backbone of Hyper-YOLO: To enhance the feature extraction ability of the base network, a Mixed Aggregation Network (MANet) is designed, as Figure 3 shown. Utilize the proposed MANet as its core computing module, thus enhancing the feature discrimination ability of the traditional C2f module in YOLO. Different from the traditional YOLO architecture, Hyper-YOLO absorbs the set of five main feature sets {B1, B2, B3, B4, B5}. In the novel approach, the neck of Hyper-YOLO is based on hypergraph computing theory, integrating the cross-level and cross-position information of these five-tuple feature sets, and finally generating the final semantic features {N3, N4, N5} at three different scales. These hierarchical semantic features are then used for the final object detection task.

[0058] The MANet architecture synergistically fuses three typical convolutional variants: 1×1 bypass convolution for channel feature recalibration, depthwise separable convolution for efficient spatial feature processing, and the C2f module for enhancing feature hierarchical integration. This fusion generates a more diverse and rich gradient flow during the training phase, thus significantly enhancing the semantic depth contained in the base features in five key stages. MANet can be expressed as follows: (12) where has 2c channels, while each of , ,..., has c channels for each feature. Finally, the semantic information of the three types of features is fused and compressed through concatenation operations and 1×1 convolution to generate with 2c channels.

[0059] The core innovation of MANet lies in the synergistic fusion of three convolutional variants, respectively for channel feature recalibration, efficient spatial feature processing, and multi-scale feature integration. These characteristics are applied to the recognition of vascular stenosis lesions, manifested in: The 1×1 bypass convolution preferentially activates features related to narrow lesions (such as changes in blood vessel wall thickness and abnormal local blood flow signals) by dynamically adjusting channel weights, and suppresses redundant background information. Vascular stenosis lesions usually present as local lumen stenosis or plaque deposition, and are easily interfered by the background in low-contrast images. The 1×1 convolution can effectively enhance the feature response in the lesion area through channel dimension screening.

[0060] DSConv decomposes the traditional convolution into depthwise convolution and pointwise convolution, reducing the number of parameters to about 1 / 8 of the traditional convolution while maintaining sensitivity to local textures. Vascular stenosis is usually accompanied by subtle morphological changes (such as sudden changes in pipe diameter and irregular plaque edges). DSConv extracts local spatial features (such as edge gradients and texture continuity) through depthwise convolution, and then fuses cross-channel semantics through pointwise convolution, balancing computational efficiency and detection accuracy.

[0061] The C2f module dynamically fuses shallow details (such as blood vessel branch morphology) and deep semantics (such as plaque distribution patterns) through feature splitting and multi-level convolution (ConvNeck module). Vascular stenosis may present in various forms such as focal and diffuse, and requires joint modeling of multi-scale features. The progressive feature fusion mechanism of the C2f module can capture both local stenosis details (such as plaque calcification) and the global blood vessel orientation.

[0062] To comprehensively fuse cross-level and cross-location information from the backbone, this embodiment further proposes a hypergraph-based cross-level and cross-location representation network (HyperC2Net). Specifically, it includes: 1. Hypergraph construction: The backbone is divided into five discrete stages. To utilize hypergraph computing to clarify the complex high-order relationships between basic features, this embodiment initiates the process by performing channel concatenation of five basic features, thereby synthesizing cross-level visual features. The hypergraph G = {V, E} is usually defined by its vertex set V and hyperedge set E. This embodiment deconstructs grid-based visual features to form the vertex set V of the hypergraph. To model the neighborhood relationships within the semantic space, a distance threshold is used to construct an ϵ-ball from each feature point, and this ball will be used as a hyperedge.

[0063] 2. Hypergraph convolution: To facilitate the transmission of high-order messages on the hypergraph structure, typical spatial-domain hypergraph convolution and additional residual connections are used for high-order learning of vertex features, as follows: , (13) where and are two neighborhood indicator functions. are trainable parameters.

[0064] 3. An example of the HGC-SCS framework: By combining the previously defined hypergraph construction and convolution strategies, a simplified example of the HGC-SCS framework, called the Hypergraph-based Cross-level and Cross-location Representation Network (HyperC2Net), is introduced, and its overall definition is as follows: , (14) where ·||· represents the matrix concatenation operation, and φ is the fusion function. In HyperC2Net, it essentially contains cross-layer information because it is the fusion of backbone features from multiple levels. Additionally, by deconstructing the grid features into a set of feature points within the semantic space and constructing hyperedges based on distances, the method of this embodiment allows for the transmission of high-order messages between vertices at different positions within the point set. This function helps capture cross-location information and enriches the model's understanding of the semantic space.

[0065] The originality of HyperC2Net lies in its use of hypergraph computing to capture the potential complex high-order correlations in the feature map. Hypergraph convolution in the semantic domain facilitates the information flow without being restricted by the grid, supporting the cross-level and cross-location high-order information propagation. This mechanism breaks the limitations of the traditional grid structure, making the feature representation more nuanced and more integrated.

[0066] By introducing hypergraph computing (HyperC2Net), Hyper-YOLO can capture the potential complex high-order correlations in the feature map and achieve cross-level and cross-location information propagation. This enhanced feature extraction ability enables the model to more effectively utilize the different information extracted by the backbone network, including cross-level and cross-location information, thereby more accurately identifying and differentiating lesions of different severities.

[0067] In addition, the Hyper-YOLO uses a multi-scale prediction head for lesion detection, including the following steps: Feed the enhanced feature map into multiple prediction heads of Hyper-YOLO, and output the bounding box (location), confidence, and classification result of each prediction head.

[0068] Among them, multi-scale prediction: (1) For the feature map of each scale, use different prediction heads for the regression and classification of the bounding box (2) The predicted box B includes: the bounding box location, the confidence of the stenosis area, and the classification of the stenosis degree.

[0069] Use non-maximum suppression (NMS) to remove the redundant boxes that detect the same lesion multiple times: (1) Calculate the IOU (Intersection over Union): Calculate the IOU for all predicted boxes; (2) Suppress low-confidence bounding boxes: When the Intersection over Union (IOU) of two bounding boxes is higher than the set threshold, retain the bounding box with a higher confidence level and remove the bounding box with a lower confidence level. (3) Output the final bounding box: After removing redundant bounding boxes, retain the bounding box with the highest confidence level as the final detection result.

[0070] Map the predicted bounding boxes output by the model from the scaled feature map back to the original image space.

[0071] In step S4, visualize the recognition results, including the annotation of the stenosis location and severity, providing a quantitative basis for clinical diagnosis. Based on step S3, by overlaying and displaying the predicted bounding boxes and classification results of vascular stenosis on the original medical image, show the stenosis location and severity.

[0072] In summary, the coronary artery stenosis dynamic quantification and recognition method based on the diffusion model and YOLO hypergraph optimization in this embodiment introduces a diffusion-based adversarial representation learning (DARL) model. Through the combination of the diffusion module and the generation module, precise segmentation of blood vessels is achieved without a large amount of labeled data. In the generation module of the DARL model, the Transformer structure is adopted to improve the image restoration quality and enhance the capture of global vascular features, especially long-range dependencies. The Hyper-YOLO algorithm with local enhanced attention is also proposed. By introducing spatial attention and channel attention modules, the recognition of vascular stenosis areas is strengthened, and the sensitivity of the model to vascular lesions is improved. In the object detection part of Hyper-YOLO, combined with the vascular segmentation mask information, the network is guided to focus on the vascular area, so as to more accurately identify the stenosis lesions in the blood vessels; detect vascular stenosis areas of different sizes through a multi-scale prediction head to achieve multi-object detection.

[0073] Based on the same concept, this embodiment also provides a coronary artery stenosis dynamic quantification and recognition system based on the diffusion model and YOLO hypergraph optimization. Please refer to Figure 4 , and this system includes: Image preprocessing module: Used to preprocess the angiography image containing noise. Through the diffusion module, learn the denoising process, gradually restore from the noisy image to the original image, and extract the background signal. Vascular segmentation module: Adopt a diffusion adversarial representation learning model based on Transformer. For the background signal, through the generation module and adversarial training, generate a vascular segmentation mask to achieve precise segmentation of blood vessels. Lesion recognition module: Adopt the Hyper-YOLO algorithm with local enhanced attention. By inserting an attention module after the convolutional layer, enhance the feature response to the vascular stenosis area, and combine the hypergraph calculation HyperC2Net to model the high-order correlations in the feature map to achieve precise recognition of vascular stenosis lesions. Result output module: Visualize the recognition results, including the annotation of the stenosis location and severity, providing a quantitative basis for clinical diagnosis.

[0074] The functions and implementation methods of the above image preprocessing module, vascular segmentation module, lesion recognition module, and result output module are as described in the above coronary artery stenosis dynamic quantification recognition method based on the diffusion model and YOLO hypergraph optimization, and will not be elaborated here.

[0075] The following selects two case samples to illustrate the specific implementation and testing: Case 1: Select the digital angiography sequence of a 58-year-old male coronary heart disease patient (BMI 28.7, suspected proximal LAD stenosis > 70%) admitted to a certain tertiary hospital as the test sample. The coronary angiography data collected by a 128-slice dual-source CT (slice thickness 0.6 mm, matrix 512×512, tube voltage 120 kV) is input into this system after being preprocessed in DICOM format. The key parameter configurations in the implementation process are as follows: The DARL model uses a 50-layer Vision Transformer architecture (patch size 4×4, hidden layer dimension 768, number of attention heads 12), the diffusion step T = 1000, and the noise scheduling parameter Adopts the cosine annealing strategy; The Hyper-YOLO backbone network is an improved DarkNet-53 (the proportion of depthwise separable convolutions is 40%), the local attention module sets the window size of k = 7, the grouping convolution dimensionality reduction coefficient γ = 8, and the sampling point density of the dynamic radius perception unit is 32 points / circumference.

[0076] In the specific implementation, first preprocess the 12th-frame diastolic angiography image of the patient (the original HU value range is [-100, 500]): Eliminate the cardiac pulsation interference through the motion artifact correction algorithm based on phase symmetry, and use non-local mean filtering (search window 21×21, similarity window 7×7) to reduce quantum noise. The preprocessed image is input into the DARL model for vascular segmentation. Among them, the diffusion module has achieved stable separation of background signals (PSNR > 38 dB) at t = 200 steps, and the generation module synchronously outputs the vascular mask (resolution 1024×1024) and the synthetic angiography image through the switchable SPADE layer. After quantitative evaluation, the Dice coefficient of the left anterior descending artery (LAD) segmentation in this case reaches 0.913 (compared with the gold standard manual segmentation), and the recall rate of small branches (diameter 0.8 - 1.2 mm) is 91.7%, which is 35.2 percentage points higher than that of the traditional U-Net.

[0077] The segmentation results are input into the Hyper-YOLO lesion recognition module after topological correction. Among them, the Mixed Aggregation Network (MANet) activates the local attention mechanism at the feature map (size 64×64×256) in the B3 stage, and generates a heat map (confidence level 0.94) for the proximal lesion area of the LAD. The dynamic radius perception unit performs vascular cross-section analysis at intervals of 0.5 mm, and detects 3 stenosis lesions (coordinate positioning error <0.3 mm). The maximum stenosis is located at 7.2 mm in the proximal segment of the LAD, and the three-dimensional quantified stenosis rate is 73.5%±2.1% (IVUS verification value 71.8%). The system completes the entire process from raw image input to lesion grading output within 1.2 seconds, with the calculation efficiency being 8 times higher than that of traditional FFR. Moreover, two vulnerable plaques are successfully identified (low-density lipoprotein core occupancy >40%, fibrous cap thickness <65 μm), and the results are highly consistent with the OCT examination conclusions (kappa coefficient 0.87).

[0078] This embodiment verifies the effectiveness of the technical solution in complex clinical scenarios: under the condition of only using 30% labeled data (the training set of this case contains 2150 unpaired angiography / background images), the DARL model accurately separates blood vessels and background noise through adversarial representation learning; Hyper-YOLO effectively models the topological association of blood vessel morphology with the help of hypergraph convolution (neighborhood construction strategy with ε = 0.35 in HyperC2Net). Its lesion grading network (SSGN) adopts a 5-layer MLP structure (hidden layer dimensions [256, 128, 64, 32, 3]), and the final accuracy of the three-level classification of stenosis degree (mild <50%, moderate 50-70%, severe >70%) reaches 96.3%. The complete processing log and parameter interaction data of this case have been stored in the hospital PACS system (storage path: / NAS / CardioAI / Case_20250218_58M_LAD / ), providing a traceable quantitative basis for subsequent clinical decisions.

[0079] Example 2: The digital angiography sequence of a 59-year-old male coronary heart disease patient admitted to a certain tertiary hospital is selected as the test sample.

[0080] First, enter the image preprocessing stage. A diffusion module based on deep learning is adopted, and the learning rate is set to 0.001 for 100 rounds of training. This module gradually restores the original image from the noisy image and improves the image quality. Before processing, the peak signal-to-noise ratio (PSNR) of the image is 22 dB, and the structural similarity index (SSIM) is 0.63. After processing, the PSNR is increased to 32 dB, and the SSIM reaches 0.88. The blood vessel contours in the image are clearer, and the noise is greatly reduced.

[0081] Next, introduce the Transformer structure into the DARL generation module, with the number of layers set to 4, the number of heads to 8, and the hidden dimension to 512, so as to improve the denoising quality and image restoration effect and capture the long-range dependence structure of blood vessels. The Dice coefficient of blood vessel segmentation before improvement was 0.72, and it was improved to 0.89 after improvement, significantly improving the accuracy of blood vessel segmentation and clearly presenting complex blood vessel branches and long-range connections.

[0082] After that, adversarial training is carried out. The generation module receives the background features of the diffusion module. The generator and discriminator perform 50 rounds of adversarial learning with a learning rate of 0.0001 for the generator and 0.0005 for the discriminator to generate realistic synthetic blood vessel images and estimated blood vessel segmentation masks. The cosine similarity between the generated image and the real image in the semantic feature space before training was 0.65, and it was improved to 0.92 after training. The model's ability to capture the semantic information of blood vessels is enhanced, and it can better distinguish blood vessel tissues from lesion features.

[0083] Then, introduce an attention module after the convolutional layer of the Hyper - YOLO network, with the compression ratio set to 16. Insert the attention module after the feature extraction layer to enhance the feature response of the key regions of blood vessels, and insert spatial attention before the prediction branch to improve the sensitivity of locating narrow regions. The results show that the feature response intensity of the key regions of blood vessels is increased by 2.2 times, and the sensitivity of locating narrow regions is increased by 35%, making the features of narrow regions more prominent.

[0084] Finally, the object detection part of Hyper - YOLO starts lesion recognition. The backbone network uses ResNet50 to extract multi-scale blood vessel features, and the attention module suppresses background noise and strengthens the edge and texture features of narrow regions. Detect different sizes of narrow lesion regions through multi-scale prediction heads at 3 scales, and output the position of the prediction box, confidence, and classification results of the narrow degree. Filter overlapping boxes through non-maximum suppression (NMS), retain high-confidence results, and map them back to the original image for overlay display. The mean average precision (mAP) of the model reaches 0.93, and the recall rate is 0.87. Two blood vessel stenoses are accurately identified in the patient's image, one with mild stenosis (28% stenosis degree) and one with moderate stenosis (52% stenosis degree).

[0085] Based on the test results, the doctor developed a treatment plan. For mildly stenotic sites, drug treatment was adopted, prescribing antiplatelet drug aspirin at 100 mg per day and statin drug atorvastatin at 20 mg per day to prevent thrombosis and stabilize plaques. For moderately stenotic sites, considering the frequent occurrence of the patient's symptoms, percutaneous coronary intervention (PCI) was decided to be performed. Under local anesthesia, a guide wire was sent along the blood vessel to the stenotic site through puncturing the patient's right femoral artery, and then a balloon catheter was inserted for dilation to restore patency of the stenotic blood vessel. Then a drug-eluting stent was implanted to keep the blood vessel open. After the operation, drug treatment was continued, and the patient was required to have regular follow-up visits to monitor the blood vessel condition and drug side effects.

[0086] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, provided that these changes fall within the scope of the claims of the present invention and its equivalent technologies, they still fall within the protection scope of the present invention.

Claims

1. A dynamic quantitative recognition method for coronary artery stenosis based on diffusion model and YOLO hypergraph optimization, characterized in that, Including: Preprocess the angiography image with noise. Through the diffusion module, learn the denoising process, gradually restore from the noisy image to the original image, and extract the background signal; Adopt a diffusion adversarial representation learning model based on Transformer. For the background signal, through the generation module and adversarial training, generate a vascular segmentation mask to achieve accurate segmentation of blood vessels; Adopt the Hyper-YOLO algorithm with local enhanced attention. By inserting an attention module after the convolutional layer, enhance the feature response to the vascular stenosis area, and combine hypergraph computing to model the high-order correlations in the feature map to achieve accurate identification of vascular stenosis lesions; Visualize the recognition results, including the annotation of the stenosis location and severity, providing a quantitative basis for clinical diagnosis.

2. The coronary artery stenosis dynamic quantification and recognition method based on diffusion model and YOLO hypergraph optimization according to claim 1, characterized in that, The preprocessing of the angiography image with noise includes: Based on the denoising diffusion probability model, gradually reduce the noise through multiple time steps. Starting from the image with the highest noise level, gradually estimate the denoised image at each intermediate step until it is restored to the original image; During the denoising process, gradually identify and restore the background signal in the image to enhance the ability to recognize the vascular morphology.

3. The coronary artery stenosis dynamic quantization recognition method based on diffusion model and YOLO hypergraph optimization according to claim 1, characterized in that, The diffusion adversarial representation learning model based on Transformer includes: Diffusion module: used to estimate the latent features, gradually remove the noise through the reverse diffusion process, and learn the background signal of the image; Generation module: introduce the Transformer structure to enhance global feature learning, capture the long-range dependencies of blood vessels, and generate a vascular segmentation mask and a synthetic angiography; Adversarial training unit: through the adversarial learning of the generator and discriminator, generate realistic vascular images and segmentation masks, and improve the model's ability to capture vascular semantic information.

4. The coronary artery stenosis dynamic quantification recognition method based on diffusion model and YOLO hypergraph optimization according to claim 3, characterized in that, Introduce the Transformer encoder module into the encoder of the generation module, and send the preliminarily extracted features of the input image into the Transformer encoder module for global context modeling; Introduce the Transformer decoder module into the decoder, process the features output by the encoder through the Transformer decoder, and gradually restore the image through upsampling and convolutional layers.

5. The coronary artery stenosis dynamic quantification recognition method based on diffusion model and YOLO hypergraph optimization according to claim 3, characterized in that Adopt multiple residual blocks with SPADE layers to form the generation module. The SPADE layer normalizes the feature map in different ways according to whether the input data includes a segmentation mask; The generation module generates a vascular segmentation mask without using the SPADE layer according to the input noisy angiography and latent features; And / or according to the input segmentation mask, use the noisy background and its latent features to synthesize the angiography using the SPADE layer.

6. The coronary artery stenosis dynamic quantification and recognition method based on diffusion model and YOLO hypergraph optimization according to claim 3, characterized in that The diffusion adversarial representation learning model based on Transformer constructs a generator to receive the noisy latent representations at different stages during the diffusion process as input and output the generated latent representations; the generated latent representations are close to the original latent representations; Construct a discriminator to distinguish between the latent representations generated by the generator and the real original latent representations; Through the way of adversarial training, alternately update the parameters of the generator and discriminator to make them reach a balance.

7. The coronary artery stenosis dynamic quantification recognition method based on diffusion model and YOLO hypergraph optimization according to claim 1, characterized in that The locally enhanced attention Hyper-YOLO algorithm includes: Insert a local attention module after the convolutional layer of Hyper-YOLO to enhance the feature response to the key regions of blood vessels and improve the localization sensitivity to narrow regions; Based on the hypergraph convolution HyperC2Net, model the complex high-order associations in the feature map, realize the information propagation across levels and positions, and improve the lesion recognition accuracy; Combine the blood vessel segmentation mask information, and detect narrow lesion regions of different sizes through a multi-scale prediction head to achieve multi-object detection.

8. The coronary artery stenosis dynamic quantification recognition method based on diffusion model and YOLO hypergraph optimization according to claim 7, wherein Inserting a local attention module after the convolutional layer of Hyper-YOLO further includes: Insert a local attention module after the key convolutional layers of the backbone network and the feature pyramid of Hyper-YOLO; The local attention module receives the feature map containing the key regions of blood vessels output by the convolutional layer, divides the feature map into k×k local windows, and independently calculates the attention weights for each window; Multiply the attention weights and the feature map element by element to obtain the enhanced feature map.

9. The coronary artery stenosis dynamic quantification recognition method based on diffusion model and YOLO hypergraph optimization according to claim 8, wherein, Detecting narrow lesion regions of different sizes through a multi-scale prediction head further includes: Input the enhanced feature map into multiple prediction heads of Hyper-YOLO, and output the bounding boxes, confidence levels, and classification results of each prediction head; Calculate the intersection over union (IoU) for all bounding boxes. When the IoU of two bounding boxes is higher than the threshold, retain the bounding box with the higher confidence level; after pairwise comparison, remove the redundant boxes and retain the bounding box with the highest confidence level as the output result; Map the output bounding boxes from the scaled feature map back to the original image space.

10. A coronary artery stenosis dynamic quantification and recognition system based on diffusion model and YOLO hypergraph optimization, characterized in that, Includes: Image preprocessing module: used to preprocess the noisy angiography image, learn the denoising process through the diffusion module, gradually recover from the noisy image to the original image, and extract the background signal; Blood vessel segmentation module: adopt a diffusion adversarial representation learning model based on Transformer, for the background signal, generate a blood vessel segmentation mask through the generation module and adversarial training to achieve accurate segmentation of blood vessels; Lesion recognition module: adopt the locally enhanced attention Hyper-YOLO algorithm, enhance the feature response to the narrow regions of blood vessels by inserting an attention module after the convolutional layer, and combine the hypergraph calculation HyperC2Net to model the high-order associations in the feature map to achieve accurate recognition of blood vessel narrow lesions; Result output module: visually display the recognition results, including the annotation of the narrow location and severity, and provide a quantitative basis for clinical diagnosis.

Images