System for automatically recognizing coronary artery dissection in DSA images on basis of u-KAN and anatomical a priori information

By combining KANs and anatomical prior information to optimize the U-Net network, the problem of insufficient accuracy in identifying coronary artery dissection in DSA images is solved, achieving efficient and accurate automatic identification of coronary artery dissection, which is suitable for clinical environments.

WO2026237167A1PCT designated stage Publication Date: 2026-11-19JIANGSU PROVINCE HOSPITAL (THE FIRST AFFILIATED HOSPITAL OF NANJING MEDICAL UNIVERSITY)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/110479
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-13
Filing Date
2025-07-25
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing deep learning methods neglect anatomical prior information when processing coronary artery dissection identification in DSA images, resulting in insufficient recognition accuracy and robustness, especially with a high misdiagnosis rate under noisy and blurred image conditions.

Method used

By combining Kolmogorov-Arnold Networks (KANs) to optimize the U-Net network structure and incorporating anatomical prior information, the accuracy of automatic identification of coronary artery dissection is improved through data preprocessing, anatomical feature extraction, model training, and result evaluation.

Benefits of technology

It significantly improves the accuracy of automatic identification of coronary artery dissection, reduces the misdiagnosis rate, and enhances the robustness and computational efficiency of the model, making it suitable for clinical applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025110479_19112026_PF_FP_ABST
    Figure CN2025110479_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed in the present invention is a system for automatically recognizing coronary artery dissection in DSA images on the basis of U-KAN and anatomical a priori information. The system comprises: a data preprocessing module, an anatomical a priori information extraction module, a U-KAN model construction module, a model training module, a dissection region automatic recognition module, and a result processing and evaluation module. By introducing a KAN layer, the present invention significantly enhances the representation capability and recognition accuracy of a U-Net model with respect to complex medical images (e.g., DSA images). The present invention innovatively incorporates anatomical a priori information (e.g., the geometric morphology of coronary arteries and vascular branches) into a deep learning model, thereby improving the sensitivity of the model to lesion regions. By means of an optimized U-KAN, the present invention achieves high computational efficiency while ensuring the recognition accuracy. The U-KAN model provided in the present invention has strong extensibility, and in addition to being applicable to the automatic recognition of coronary artery dissection, the model can also be extended to other types of medical image analysis tasks including the automatic detection and segmentation of brain CT images and tumor MRI images.
Need to check novelty before this filing date? Find Prior Art

Description

An Automatic Coronary Artery Dissection Identification System Based on U-KAN and Anatomical Priors using DSA Images Technical Field

[0001] This invention belongs to the fields of medical image processing and computer vision, and specifically relates to an autonomous identification method and system for coronary artery dissection in DSA images based on Kolmogorov-Arnold UNet and anatomical priors. Background Technology

[0002] Coronary artery dissection (CAD) is a serious coronary vascular disease, usually caused by the rupture of the vessel wall, leading to obstructed blood flow and potentially causing myocardial ischemia, myocardial damage, or even cardiac arrest. Early diagnosis of CAD is crucial for patient prognosis; timely intervention and treatment can significantly improve survival rates and quality of life. Digital subtraction angiography (DSA), as an important imaging technique, has been widely used in the diagnosis of coronary artery disease. DSA can clearly display the vascular structure, helping doctors determine the presence and severity of lesions. However, because DSA images are often affected by noise, blurring, and background information, the manual analysis process is often very complex and easily affected by the operator's experience and visual acuity, resulting in low diagnostic accuracy and a high rate of misdiagnosis.

[0003] In recent years, deep learning, especially convolutional neural networks (CNNs), has become an important tool in the field of medical image analysis. The superior performance of CNNs in tasks such as image segmentation and object detection has made them the mainstream method for automated medical image analysis. Deep learning technology has demonstrated enormous potential, particularly in the automatic identification and segmentation of coronary artery dissection. Traditional deep learning methods, especially the U-Net model, have been widely applied to medical image segmentation tasks and have achieved good results. However, existing deep learning methods still have some limitations, especially when dealing with complex and noisy medical images such as coronary artery dissection. Traditional models often neglect the role of prior anatomical information, leading to insufficient recognition accuracy and robustness.

[0004] Anatomical prior information, particularly the geometry of coronary arteries, vascular distribution, and their relationship with surrounding structures, often plays a crucial role in improving the accuracy of coronary artery dissection identification. However, traditional image analysis methods and deep learning networks rarely explicitly incorporate anatomical information into the model's learning process, leading to a high misdiagnosis rate. Therefore, effectively combining anatomical prior information and deep learning techniques to enhance the automatic identification capability of coronary artery dissection in DSA images remains a pressing technical challenge.

[0005] In recent years, novel neural network architectures, especially Kolmogorov-Arnold Networks (KANs), have made significant progress in improving model accuracy and interpretability. KANs, by introducing nonlinear learnable activation functions derived from the Kolmogorov-Arnold representation theorem, have revolutionized the learning process of traditional neural networks. This method not only enhances the network's expressive power but also strengthens the model's stability and robustness in complex tasks. Particularly in the field of medical image segmentation, KANs have demonstrated their advantages in processing complex image features, better capturing subtle lesion features and improving segmentation accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic method and system for identifying coronary artery dissection in digital subtraction angiography (DSA) images based on Kolmogorov-Arnold UNet (U-KAN) and anatomical prior information. By introducing Kolmogorov-Arnold Networks (KANs) to optimize the U-Net network structure and combining it with anatomical prior information, the automatic identification accuracy of coronary artery dissection can be significantly improved, and some limitations of traditional deep learning models in medical image analysis, such as noise interference, image blurring, and loss of detailed features, can be effectively overcome.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an automatic identification system for coronary artery dissection based on U-KAN and anatomical priors in DSA images, comprising:

[0008] The data preprocessing module is used to perform denoising, enhancement, and standardization on DSA images;

[0009] The anatomical prior information extraction module is used to extract the anatomical features of the coronary arteries from the preprocessed image as prior features, and to extract anatomical structure data from the medical knowledge base or manually annotated as prior information.

[0010] The U-KAN model building module is used to build the U-KAN model by optimizing the U-Net architecture and introducing a KAN layer in the U-Net encoder part.

[0011] The model training module is used to perform supervised learning training on the U-KAN model, using the extracted prior features and prior information as the dataset.

[0012] The automatic identification module for dissecting regions is used to mark the coronary artery dissecting regions in new DSA images using a trained U-KAN model.

[0013] The results processing and evaluation module is used to optimize the segmentation of the labeled results output by the model, evaluate the results recognized by the model, and verify the performance of the model.

[0014] Furthermore, the denoising process for DSA images employs image denoising algorithms and deep learning methods to reduce noise and blur in the images. For artifacts and motion artifacts that may exist in the images, artifact removal techniques are used for optimization. The enhancement operation adjusts the brightness, contrast, and saturation of the images. The standardization process unifies the size, resolution, and color channels of all input images.

[0015] Furthermore, the anatomical features include the geometry of the coronary arteries, their branching structures, and their centerlines. The extraction of these anatomical features employs a centerline extraction algorithm.

[0016] Further anatomical data include the marginal features of blood vessels and the barbed anatomical features of coronary artery dissection.

[0017] Furthermore, the extraction of the edge features of blood vessels is performed using the following steps:

[0018] Use edge detection algorithms to extract the edges of blood vessels from DSA images;

[0019] Edge refinement is achieved through a deep learning-based edge detection network to improve the accuracy of blood vessel edge detection.

[0020] Furthermore, the extraction of the barbed anatomical features of coronary artery dissection was performed using the following steps:

[0021] Morphological analysis methods were used to analyze the images and identify the barbed structure;

[0022] By combining deep learning methods, the barbed features of the interlayer region are automatically learned and extracted through a convolutional neural network.

[0023] Furthermore, the U-KAN network structure constructed by the U-KAN model building module includes:

[0024] The encoder section is used to extract features from the input image. It uses a convolutional neural network to extract low-level features of the image, including edges, textures, and colors, and reduces the spatial dimension of the image through downsampling. Each layer in the encoder uses a KAN layer instead of a traditional activation function, and enhances the model's ability to learn complex image features by introducing a non-linear learnable activation function.

[0025] Decoder section: Used to restore the spatial resolution of the image through inverse convolution, reconstruct the details of the image through upsampling, and perform fine reconstruction of the boundaries and shapes of the interlayer region;

[0026] Output layer: Used to output the classification probability of each pixel through the Softmax layer, and obtain the binarized segmentation result, representing the region of coronary artery dissection in the image.

[0027] Furthermore, the model training module uses a dataset that includes images with coronary artery dissection annotations and normal images. It employs the cross-entropy loss function to measure the difference between the model output and the true labels, and uses the Dice coefficient or IoU to optimize network training for the segmentation task. The model training module uses data augmentation techniques to generate diverse training samples to reduce overfitting. Simultaneously, it employs an adaptive learning rate adjustment strategy, using a larger learning rate in the early stages of training for rapid convergence, and gradually decreasing the learning rate in later stages to improve training accuracy. After each certain number of training epochs, the model training module uses a validation set to evaluate the model, monitoring its performance on new data to ensure that there is no overfitting during training.

[0028] Furthermore, the result output by the interlayer region automatic recognition module is a binary image, in which the interlayer region is marked as a highlighted area.

[0029] Furthermore, the result processing and evaluation module optimizes the segmentation of the labeled results output by the model by refining the boundaries and removing noise, and using morphological operations to optimize the segmentation results. The evaluation of the model recognition results verifies the model's performance by calculating the Dice coefficient, accuracy, recall, and F1-score, and comparing them with the results of manual annotation to verify the model's accuracy and reliability.

[0030] The technical solution of the present invention includes the following specific steps:

[0031] Data preprocessing:

[0032] First, denoising is performed on the DSA images. Traditional image denoising algorithms (such as Gaussian filtering and mean filtering) and deep learning methods (such as convolutional neural network denoising) are used to reduce noise and blur in the images. In particular, specialized artifact removal techniques are employed to optimize the images for potential artifacts and motion artifacts.

[0033] Secondly, image enhancement operations are performed, including adjustments to brightness, contrast, saturation, etc., so that the model can better learn the key features in the image during training.

[0034] Finally, image standardization is performed to unify the size, resolution, and color channels of all input images, ensuring the consistency of the data input to the network.

[0035] Extraction of prior anatomical information:

[0036] After image preprocessing, anatomical features of the coronary arteries are extracted using image analysis techniques (such as edge detection and image segmentation). For example, a vessel centerline extraction algorithm is used to extract the geometric shape and branching information of the coronary arteries, which are then input into a deep learning model as prior features.

[0037] At the same time, by combining medical knowledge base or manually annotated anatomical structure data, the morphology, distribution pattern of blood vessels and dissection characteristics (such as intimal dissection, hematoma formation, etc.) are used as prior information to help the model better understand the lesion area in the image.

[0038] U-KAN model construction:

[0039] The core of this invention is to improve recognition accuracy by introducing Kolmogorov-Arnold Networks (KANs) into an optimized U-Net architecture. Traditional U-Net networks employ an encoder-decoder structure, where the encoder extracts features from the image, and the decoder reconstructs details. To enhance the model's expressive power, a KAN layer is introduced into the encoder part of U-Net. Kolmogorov-Arnold Networks (KANs): KANs enhance the network's expressive power through non-linear, learnable activation functions, enabling the network to more accurately fit complex image features. Specifically, the introduction of KANs allows for higher-dimensional modeling of complex non-linear relationships in the image within the encoder layer, further improving the network's ability to identify blood vessels and dissection features.

[0040] In the decoder part of U-Net, the KAN layer helps the network better restore image details, especially the boundary and morphology of the coronary artery dissection region, through an improved activation function, reducing the boundary blurring and over-smoothing problems in the traditional U-Net model.

[0041] Model training:

[0042] The U-KAN model was trained using supervised learning with a pre-labeled DSA image dataset, including images with coronary artery dissection annotations and normal images. During training, the model parameters were updated by calculating the difference between the model output and the true labels (e.g., using the cross-entropy loss function).

[0043] During training, data augmentation techniques (such as rotation, flipping, scaling, and cropping) are used to generate diverse training samples, increasing the model's generalization ability and reducing overfitting.

[0044] Meanwhile, an adaptive learning rate adjustment strategy is adopted, which uses a larger learning rate in the early stage of training to achieve rapid convergence, and gradually reduces the learning rate in the later stage to improve training accuracy.

[0045] Automatic identification of mezzanine areas:

[0046] The trained U-KAN model is used to infer new DSA images. The model extracts, encodes, and decodes features from the input image to generate a classification result for each pixel in the image, thus marking the coronary artery dissection region in the image.

[0047] This process learns from features such as vascular morphology, intimal dissection, and hematoma to accurately identify and segment the dissecting region. The output is a binary image, in which the dissecting region is marked as a highlighted area.

[0048] Post-processing and evaluation of results:

[0049] The segmentation results output by the model are post-processed. First, noise in the image is filtered out, and then the boundaries of the interlayer regions are refined to ensure the accuracy of the boundaries. The segmentation results can be further optimized through morphological operations (such as erosion and dilation).

[0050] Finally, the model recognition results are evaluated. Commonly used evaluation metrics (such as Dice coefficient, accuracy, recall, etc.) are used to quantify the recognition accuracy, and the results are compared with those of manual annotation to verify the model's performance.

[0051] The advantages of this invention are as follows:

[0052] 1. Introduction of Kolmogorov-Arnold Networks (KANs) to optimize U-Net: This invention significantly enhances the U-Net model's expressive power and recognition accuracy for complex medical images (such as DSA images) by introducing KAN layers. The nonlinear activation function of KAN can capture more complex image features, effectively improving the recognition ability of coronary artery dissection regions, especially in images with high noise interference.

[0053] 2. Integrating Anatomical Prior Information: This invention innovatively incorporates anatomical prior information (such as the geometry of coronary arteries and vascular branches) into the deep learning model, thereby improving the model's sensitivity to lesion areas. The introduction of anatomical prior information not only helps the network better understand vascular structures but also improves the accurate identification of coronary artery dissection areas, reducing false positives and false negatives. Simultaneously, it boasts high computational efficiency. Compared to traditional deep learning models, the U-KAN network can achieve high-precision recognition with lower computational costs during image processing, making it suitable for widespread application in clinical settings.

[0054] 4. Multi-task learning and scalability: The U-KAN model proposed in this invention has strong scalability. In addition to being applied to the automatic identification of coronary artery dissection, it can also be extended to other types of medical image analysis tasks, such as the automatic detection and segmentation of brain CT and tumor MRI images. Attached Figure Description

[0055] Figure 1 is a schematic diagram of the UKAN network according to an embodiment of the present invention.

[0056] Figure 2 is a flowchart of the system integration into clinical decision-making according to an embodiment of the present invention. Detailed Implementation

[0057] To better understand the specific embodiments of the present invention, the following describes in detail, with reference to the accompanying drawings and actual operation, the automatic identification method for coronary artery dissection based on digital subtraction angiography (DSA) images of the present invention.

[0058] 1. System Architecture

[0059] The system architecture of this invention mainly consists of the following modules:

[0060] Data acquisition module: Used to acquire DSA image data. Images can be acquired through conventional DSA equipment or from public databases (such as the Radiology Database). The image format is a standard medical imaging format, such as DICOM (Digital Imaging and Communications in Medicine) format.

[0061] Data preprocessing module: Denoising, enhancing, and normalizing the acquired DSA images. Denoising uses deep denoising networks (such as convolutional autoencoders) to remove image noise; enhancing includes contrast and brightness adjustments to improve image detail visibility; normalization ensures all images are the same size, facilitating subsequent input into the neural network.

[0062] Anatomical Information Extraction Module: This module uses traditional image processing methods and deep learning models (such as deep neural networks based on image segmentation) to extract anatomical features from DSA images, including the geometry of blood vessels, branching structures, and the centerline of blood vessels. This anatomical information is input as prior features into the deep learning model to enhance the model's semantic understanding.

[0063] Deep Learning Model (U-KAN) Module: This module is the core of this invention, including an optimized U-Net network combined with Kolmogorov-Arnold Networks (KAN). The U-KAN structure includes an encoder part, a decoder part, and KAN layers, which automatically identify and segment coronary artery dissection regions by combining anatomical prior information.

[0064] The post-processing and evaluation module optimizes the model's output, including boundary refinement, noise removal, and morphological operations. Finally, common medical image evaluation metrics (such as Dice coefficient, accuracy, and recall) are used to evaluate the recognition performance.

[0065] 2. Data Preprocessing and Augmentation

[0066] Data preprocessing is a crucial step before model training, directly impacting the model's convergence speed and final performance. The specific preprocessing steps are as follows:

[0067] Denoising: Use deep denoising networks (such as convolutional autoencoders or deep denoising convolutional neural networks) to denoise images and remove blur and artifacts caused by device noise or poor shooting conditions.

[0068] Augmentation: Use data augmentation methods (such as rotation, flipping, cropping, translation, transformation, etc.) to increase the diversity of the training dataset and avoid overfitting of the model.

[0069] Standardization: Adjust the images to a fixed size (e.g., 256×256 or 512×512) and perform grayscale standardization to scale the pixel values ​​of all images to the [0,1] range to ensure data uniformity.

[0070] 3. Extraction of prior anatomical information

[0071] In this invention, the prior anatomical information mainly includes the edge features of blood vessels and the barbed anatomical structure of coronary artery dissection. These features play a crucial role in the diagnosis of coronary artery dissection, helping deep learning models to better locate the lesion area, especially to identify the complex structure of the dissection.

[0072] The specific extraction steps are as follows:

[0073] Blood vessel edge feature extraction:

[0074] Edge detection algorithms (such as Canny edge detection) are used to extract the edges of blood vessels from DSA images. The edges of blood vessels are important indicators of vascular distribution and dissection regions, especially when dissection occurs, as the separation of the intima and adventitia will show significant changes in the edge region.

[0075] In addition, edge refinement can be achieved through deep learning edge detection networks (such as Holistically-Nested Edge Detection, HED) to improve the accuracy of blood vessel edges.

[0076] Spiral structure feature extraction:

[0077] The barbed features of coronary artery dissection typically appear as irregular protrusions or barbs at the site of separation between the intima and adventitia. These barbed features are key to identifying coronary artery dissection.

[0078] Morphological analysis methods (such as local maximum detection) are used to analyze the images and identify these barbed structures. Furthermore, deep learning methods can be combined, using convolutional neural networks (CNNs) to automatically learn and extract the barbed features of the interstitial regions. Fusion of anatomical prior information:

[0079] The extracted features of the blood vessel edges and barbed structures will be used as anatomical prior information. Combined with the pixel information in the input image, these features will be fed into the optimized U-KAN network to help the network more accurately identify the coronary artery dissection region.

[0080] Anatomical prior information can serve as an auxiliary signal, guiding the model to focus on regions with special anatomical features in images during network training, especially in the identification of interstitial regions, providing additional semantic information.

[0081] 4. Construction of U-KAN Network Model

[0082] This invention optimizes U-Net by combining it with Kolmogorov-Arnold Networks (KAN), forming the U-KAN network. The U-KAN network structure includes:

[0083] The encoder section is responsible for extracting features from the input image. It uses a convolutional neural network to extract low-level features (such as edges, textures, and colors) and reduces the spatial dimensionality of the image through downsampling. Each layer in the encoder uses a KAN layer instead of a traditional activation function (such as ReLU), introducing a non-linear learnable activation function to enhance the model's ability to learn complex image features.

[0084] Decoder section: The decoder restores the spatial resolution of the image through inverse convolution (transposed convolution). It reconstructs image details through upsampling, especially finely reconstructing the boundaries and morphology of the dissecting region. The introduction of the KAN layer allows the decoder to better capture the details of blood vessel morphology, thereby improving the segmentation effect of the coronary artery dissection region.

[0085] KAN layer: A KAN layer is added after each convolutional layer. This layer uses a non-linear activation function derived from the Kolmogorov-Arnold representation theorem to enhance the network's modeling capabilities. The KAN layer helps the model understand complex non-linear relationships and subtle image features, especially when identifying lesion areas, significantly improving the network's accuracy and robustness.

[0086] Output layer: Finally, the classification probability of each pixel is output through the Softmax layer to obtain the binarized segmentation result, which represents the region of coronary artery dissection in the image.

[0087] 5. Model Training and Optimization

[0088] During training, supervised learning is performed using a labeled DSA image dataset. Specific training steps include:

[0089] Loss function: The cross-entropy loss function is used to measure the difference between the model output and the true label. For segmentation tasks, metrics such as Dice coefficient and IoU (Intersection over Union) can be used to optimize the training of the network.

[0090] Optimization Algorithm: Gradient descent using the Adam optimizer minimizes the loss function by continuously updating network weights. Early Stopping and Regularization: Overfitting is avoided through early stopping, and Dropout, L2 regularization, and other methods are used to further improve the model's generalization ability.

[0091] Evaluation during training: After each training cycle, the model is evaluated using a validation set to monitor its performance on new data and ensure that there is no overfitting during training.

[0092] 6. Results Post-processing and Validation

[0093] After the model is trained, it is validated using a test set to evaluate its performance in identifying coronary artery dissection.

[0094] 1. Post-processing: The output of the model is refined by boundary refinement and noise removal. Morphological operations (such as dilation, erosion, opening and closing operations) are used to optimize the segmentation results to ensure that the segmentation boundary of the interlayer region is more accurate.

[0095] 2. Evaluation metrics: The performance of the model is evaluated by calculating the Dice coefficient, accuracy, recall, F1-score, etc., and compared with the results of manual annotation to verify the accuracy and reliability of the model.

[0096] The core principle of this invention lies in combining Kolmogorov-Arnold Networks (KAN) with anatomical prior information to optimize the traditional U-Net network. This enables it to more efficiently capture complex nonlinear relationships and detailed features when processing digital subtraction angiography (DSA) images, especially for the identification and segmentation of coronary artery dissections. By introducing these innovations, the U-KAN network demonstrates stronger modeling capabilities and higher accuracy in image segmentation tasks.

[0097] 1. Kolmogorov-Arnold Networks (KAN) optimizes U-Net

[0098] The traditional U-Net architecture is a classic image segmentation network that uses an encoder-decoder architecture to extract image features through convolutional operations and reconstruct image details through upsampling. However, when processing complex medical images, especially when dealing with complex lesion areas (such as coronary artery dissection), U-Net is prone to neglecting or losing detailed features, resulting in insufficient segmentation accuracy.

[0099] The introduction of KAN solves this problem. KAN's nonlinear learnable activation function effectively enhances the expressive power of neural networks, enabling them to capture more complex nonlinear patterns. Specifically, KAN reconstructs the activation function in traditional U-Net by using the nonlinear function from the Kolmogorov-Arnold representation theorem, thus enhancing its ability to model image details and complex relationships. Its advantages include:

[0100] Nonlinear modeling capability: The nonlinear activation function introduced by KAN has a stronger fitting ability and can handle complex local variations in images, especially the tortuousness of vascular structures and the barbed features of dissections. This nonlinear modeling capability can significantly improve the accuracy of capturing complex structures.

[0101] Enhanced Feature Representation: While activation functions like ReLU used in traditional U-Net are effective for learning some structures, they fall short in representing complex boundaries and details in images (such as subtle variations in sandwich structures). KAN, through its higher-order nonlinear activation function, enables the network to generate stronger responses to these features, thereby improving the accurate segmentation of sandwich regions.

[0102] Efficiently learning complex patterns: By introducing a KAN layer, the network can better learn complex patterns in images, such as changes in coronary arteries and morphological features of lesion areas. The network can adaptively adjust parameters during training, thereby more accurately identifying and segmenting coronary artery dissections.

[0103] 2. Fusion of prior anatomical information

[0104] Anatomical prior information is a crucial component of this invention, specifically referring to information about vessel margins and the barbed anatomical structure of coronary artery dissection. Traditional deep learning methods often neglect these anatomical features, relying directly on raw pixel information for training. This approach may cause the network to miss some crucial structural information in the image, particularly for complex lesion areas, potentially leading to insufficient model recognition accuracy.

[0105] To address this issue, this invention combines prior anatomical information with a deep learning model, providing stronger guidance. Specific methods include:

[0106] Guidance from vessel edge features: Vessel edges are important landmarks in coronary artery dissection. By accurately extracting vessel edges, the network can more precisely locate the vessel and the lesion area of ​​the dissection. Vessel edges typically appear as high-contrast image regions, which help the model focus on the structure of the vessel and the location of the lesion.

[0107] Guidance from the barbed features of dissection: The barbed features of coronary artery dissection typically appear at the site of separation between the intima and adventitia. These protrusions or barbs are key to diagnosing dissection. By incorporating this prior feature into the network, the model can actively focus on these detailed areas during the identification process, improving the segmentation accuracy of the dissection region.

[0108] Integration of Prior Information: In the U-KAN network, anatomical prior information is introduced through feature fusion. Specifically, the features of blood vessel edges and the barbed features of the dissection are extracted using traditional image processing methods (such as edge detection and morphological manipulation), and then fed as additional input features along with the original image into the KAN-optimized U-Net model. This information fusion enables the network to learn details related to the dissection region more effectively, improving the accuracy and reliability of segmentation.

[0109] 3. Improve recognition accuracy and stability

[0110] By combining the nonlinear activation function of KAN with anatomical prior information, the U-KAN network demonstrates significant advantages in the automatic identification and segmentation of coronary artery dissection. Specifically, this is reflected in the following aspects:

[0111] High-precision segmentation: Due to the enhanced modeling ability of complex image features by the nonlinear activation function of KAN, U-KAN can more accurately capture detailed features in images, especially in the identification of dissecting regions, effectively distinguishing the boundary between dissections and normal blood vessels. The guidance of anatomical prior information further improves the accuracy of segmentation, maintaining high accuracy even against complex image backgrounds.

[0112] Reducing misdiagnosis and missed diagnosis: Traditional methods often lead to misdiagnosis and missed diagnosis due to noise and image blurring, especially when the boundaries of the dissection area are unclear. However, the U-KAN network, by fusing vascular edge and barbed features, can better cope with noise and blurring interference, accurately locate the lesion area, and significantly reduce the misdiagnosis rate.

[0113] High computational efficiency: Although the KAN layer enhances the network's expressive power and nonlinear learning ability, U-KAN can maintain high accuracy while avoiding a significant increase in computational costs through reasonable network design and optimization. Compared with traditional methods, U-KAN has significant advantages in processing speed and computational resource consumption, making it suitable for large-scale clinical applications.

[0114] The preferred embodiments of this utility model have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of this utility model without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of this utility model through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A U-KAN and anatomic prior-based automatic identification system for coronary dissection in DSA images, characterized in that include: The data preprocessing module is used to perform denoising, enhancement, and standardization on DSA images; The anatomical prior information extraction module is used to extract the anatomical features of the coronary arteries from the preprocessed image as prior features, and to extract anatomical structure data from the medical knowledge base or manually annotated as prior information. The U-KAN model building module is used to build the U-KAN model by optimizing the U-Net architecture and introducing a KAN layer in the U-Net encoder part. The model training module is used to perform supervised learning training on the U-KAN model, using the extracted prior features and prior information as the dataset. The automatic identification module for dissecting regions is used to mark the coronary artery dissecting regions in new DSA images using a trained U-KAN model. The results processing and evaluation module is used to optimize the segmentation of the labeled results output by the model, evaluate the results recognized by the model, and verify the performance of the model.

2. The U-KAN and anatomic prior-based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The denoising process for DSA images employs image denoising algorithms and deep learning methods to reduce noise and blur in the images. For artifacts and motion artifacts that may exist in the images, artifact removal techniques are used for optimization. The enhancement operation adjusts the brightness, contrast, and saturation of the images. The standardization process unifies the size, resolution, and color channels of all input images. 3.The U-KAN and anatomical prior-based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The anatomical features include the geometry, branching structure, and centerline of the coronary arteries. The extraction of anatomical features employs a centerline extraction algorithm.

4. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The anatomical data includes the marginal features of blood vessels and the barbed anatomical features of coronary artery dissection.

5. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 4, characterized in that: The extraction of the edge features of the blood vessels is performed using the following steps: Use edge detection algorithms to extract the edges of blood vessels from DSA images; Edge refinement is achieved through a deep learning-based edge detection network to improve the accuracy of blood vessel edge detection.

6. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 4, characterized in that: The extraction of the barbed anatomical features of the coronary artery dissection was performed using the following steps: Morphological analysis methods were used to analyze the images and identify the barbed structure; By combining deep learning methods, the barbed features of the interlayer region are automatically learned and extracted through a convolutional neural network.

7. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The U-KAN network structure constructed by the U-KAN model construction module includes: The encoder section is used to extract features from the input image. It uses a convolutional neural network to extract low-level features of the image, including edges, textures, and colors, and reduces the spatial dimension of the image through downsampling. Each layer in the encoder uses a KAN layer instead of a traditional activation function, and enhances the model's ability to learn complex image features by introducing a non-linear learnable activation function. Decoder section: Used to restore the spatial resolution of the image through inverse convolution, reconstruct the details of the image through upsampling, and perform fine reconstruction of the boundaries and shapes of the interlayer region; Output layer: Used to output the classification probability of each pixel through the Softmax layer, and obtain the binarized segmentation result, representing the region of coronary artery dissection in the image.

8. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The model training module uses a dataset that includes images with coronary artery dissection annotations and normal images. It employs the cross-entropy loss function to measure the difference between the model output and the true labels. For the segmentation task, it uses the Dice coefficient or IoU to optimize network training. The module uses data augmentation techniques to generate diverse training samples to reduce overfitting. Simultaneously, it employs an adaptive learning rate adjustment strategy, using a larger learning rate in the early stages of training for rapid convergence, and gradually decreasing the learning rate in later stages to improve training accuracy. After each certain number of training epochs, the module uses a validation set to evaluate the model, monitoring its performance on new data to ensure that overfitting does not occur during training.

9. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The result output by the automatic identification module of the interlayer region is a binary image, in which the interlayer region is marked as a highlighted area.

10. The U-KAN and anatomical prior based DSA image coronary dissection automatic identification system according to claim 1, characterized in that: The result processing and evaluation module optimizes the segmentation of the labeled results output by the model by refining the boundaries and removing noise, and by using morphological operations to optimize the segmentation results. The evaluation of the model recognition results verifies the model's performance by calculating the Dice coefficient, accuracy, recall, and F1-score, and comparing the results with those of manually labeled results to verify the model's accuracy and reliability.