Coronary artery affine registration method, equipment, medium and product

Abstract

The invention discloses a coronary artery affine registration method and device, a medium and a product, and relates to the field of medical image processing, and the method comprises the steps: generating a training sample based on a standardized image; an encoder structure is adopted, and a feature extraction model comprising an encoding unit and a descriptor extraction unit is constructed by fusing a convolutional neural network, a Mama module and an attention mechanism; the coding unit extracts feature expressions of an input image on different scales by stacking a multi-layer heterogeneous dual-path module and an attention enhancement module and introducing a Leaky ReLU activation function; training a feature extraction model by adopting the training sample, and inputting an image block obtained by cutting an image to be registered into the trained feature extraction model to obtain a feature descriptor; and generating a corresponding point set based on the feature descriptors to determine an affine registration matrix to obtain a registration result. According to the method, the precision, the efficiency and the robustness of the coronary artery X-ray contrast image in an affine registration task can be improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and in particular to a method, device, medium, and product for coronary artery affine registration. Background Technology

[0002] Traditional X-ray angiography (XA) image registration techniques face numerous challenges and limitations in practical applications. Feature-based registration algorithms (such as Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), Oriented Fast and Rotated BRIEF (ORB)) often struggle to achieve accurate feature extraction and matching when dealing with medical images that are structurally complex and have high texture similarity. They are also highly sensitive to noise interference, particularly in accurately locating minute structures within the coronary arteries. Furthermore, the features relied upon by these methods are mostly manually constructed, resulting in high computational costs and slow processing speeds, failing to meet clinical requirements for registration efficiency. More critically, traditional methods are prone to getting trapped in local optima when faced with images exhibiting significant deformation, lacking robustness and struggling to handle complex spatial transformations.

[0003] The rise of deep learning technology has brought a new opportunity to solve the above problems. Leveraging its powerful feature representation capabilities, deep models can automatically learn multi-layered and complex structural information from data, eliminating reliance on manually designed features and thus significantly improving registration accuracy. Simultaneously, the end-to-end processing framework effectively reduces human intervention in intermediate steps, significantly improving overall operational efficiency. When processing images with large deformations, deep models demonstrate stronger adaptability by learning nonlinear mapping relationships. Furthermore, deep learning performs particularly well in large-scale data environments, extracting more robust feature representations. Combined with the development of computing hardware and inference acceleration technologies in recent years, it has gradually achieved the transition from high performance to clinical application.

[0004] In summary, deep learning methods, with their superior feature modeling capabilities, processing efficiency, and deformation adaptability, offer a more accurate, efficient, and robust technical approach for coronary X-ray angiography image registration, demonstrating application potential that surpasses traditional methods and providing solid support for the intelligent diagnosis and intervention of coronary artery diseases. However, most existing deep learning registration frameworks contain downsampling modules, which can easily lead to the loss of image details. Furthermore, feature maps from all channels are treated equally, and important information beneficial to the task, such as edge responses, is easily overlooked. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a coronary artery affine registration method, device, medium, and product.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for coronary artery affine registration, comprising: Obtain coronary artery X-ray angiography images and perform standardization processing on the coronary artery X-ray angiography images to obtain standardized images; Key points are extracted from the standardized image, and image patches are cropped centered on the detected key points to generate training samples; An encoder structure is adopted, and a feature extraction model is constructed by fusing convolutional neural networks, Mamba modules, and attention mechanisms. The feature extraction model includes an encoding unit and a descriptor extraction unit. The encoding unit extracts feature representations of the input image at different scales by stacking multiple heterogeneous dual-path modules and enhanced attention modules, and introducing the Leaky ReLU activation function. The descriptor extraction unit includes a feature fusion module. The feature extraction model is trained using the training samples until a set condition is met, resulting in a well-trained feature extraction model. Acquire the image to be registered and extract the key points of the image to be registered; The image to be registered is cropped into image blocks based on the key points of the image to be registered; The image patch is input into the trained feature extraction model to obtain feature descriptors; Generate a corresponding point set based on the feature descriptor; Based on the corresponding point set, the affine registration matrix is ​​determined, and the registration result is obtained.

[0007] Optionally, the heterogeneous dual-path module uses a channel partitioning strategy to divide the input features into two groups in the channel dimension, one group inputs to the CNN branch and the other group inputs to the SSM branch, fusing local features and global features; the heterogeneous dual-path module uses downsampling operation to extract information at different scales.

[0008] Optionally, the CNN branch is used to implement local receptive field modeling; the SSM branch is based on the SS2D algorithm and introduces a cross-scan module on the basis of adopting a two-dimensional selective scanning mechanism to expand the receptive field range and realize global information modeling.

[0009] Optionally, the enhanced attention module achieves the identification of coronary artery structures by fusing a learnable Sobel spatial attention mechanism with a dynamic channel-gated attention mechanism.

[0010] Optionally, the encoding unit includes a shallow stage, an intermediate stage, and a deep stage; In the shallow stage, a learnable Sobel spatial attention mechanism is integrated on the basis of the heterogeneous dual-path module to enhance the perception and modeling of local image information. In the intermediate layer stage, an enhanced attention module is introduced on the basis of the heterogeneous dual-path module to fuse channel and spatial dimension information and adopt a collaborative enhancement mechanism to achieve deep fusion of local structure and intermediate semantic features. In the deeper stage, a dynamic channel gating attention mechanism is adopted based on the heterogeneous dual-path module to suppress redundant channels.

[0011] Optionally, the SIFT algorithm is used to extract key points from the standardized image and the key points from the image to be registered.

[0012] Optionally, generating a corresponding point set based on the feature descriptor includes: Determine the Euclidean distance between the feature descriptors to obtain the similarity matrix; Initial corresponding point pairs are obtained based on the similarity matrix; The RANSAC algorithm is used to filter the initial corresponding point pairs to obtain matching point pairs, thereby generating the corresponding point set.

[0013] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the coronary artery affine registration method provided above.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the coronary artery affine registration method described above.

[0015] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the coronary artery affine registration method described above.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, device, medium, and product for coronary artery affine registration. It employs an encoder structure and constructs a feature extraction model by fusing convolutional neural networks, Mamba modules, and attention mechanisms to extract feature descriptors. By incorporating a heterogeneous dual-pathway module into the feature extraction model, local and global feature expressions can be efficiently fused, further enhancing the model's ability to model complex structures in coronary X-ray angiography images. The enhanced attention module enables efficient identification of key coronary artery structures, demonstrating significant advantages in feature enhancement and selective suppression, effectively improving the discriminative and structural sensitivity of feature expressions. The feature fusion module effectively integrates feature expressions at different scales, avoiding the neglect of important information and improving the discriminative and robust nature of feature descriptors during the matching stage. This, in turn, improves the accuracy, efficiency, and robustness of coronary X-ray angiography images in affine registration tasks, demonstrating promising clinical application prospects and widespread value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of a coronary artery affine registration method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the coronary artery affine registration process provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a feature extraction model provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an enhanced attention module provided in an embodiment of this application; Figure 5 This is a schematic diagram of the attention structure of a dynamic gating channel provided in an embodiment of this application; Figure 6 A schematic diagram of the structure of learnable Sobel spatial attention provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a heterogeneous dual-path module provided in an embodiment of this application; Figure 8 A schematic diagram of the calculation process of the SSM branch provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] In one exemplary embodiment, this application provides a coronary artery affine registration method. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is described using a server as an example. Figure 1 and Figure 2 As shown, the method includes: Step 100: Acquire coronary angiography images (hereinafter referred to as coronary artery images) and standardize them to obtain standardized images. For example, standardize the coronary angiography images to a 512×512 pixel resolution.

[0022] Step 101: Extract key points from the standardized image and crop image patches centered on the detected key points to generate training samples.

[0023] Step 102: An encoder structure is adopted, and a feature extraction model (hereinafter referred to as the model) is constructed by fusing a convolutional neural network, a Mamba module, and an attention mechanism. The feature extraction model (which may be referred to as the EGHardNet network in this application) includes an encoding unit and a descriptor extraction unit. The encoding unit extracts feature representations of the input image at different scales by stacking multiple layers of heterogeneous dual-path (Convolutional State Space Model Block, ConvSSMBlock) modules and enhanced attention (Sobel Dynamic Convolutional Block Attention Module, SDCBAM) modules, and introducing the Leaky ReLU activation function. The descriptor extraction unit includes a feature fusion module. Among them, the Mamba module is a plug-and-play visual processing module based on a state space model (SSM).

[0024] Step 103: Train the feature extraction model using training samples until the set conditions are met (e.g., the set number of training iterations are reached), and the trained feature extraction model is obtained.

[0025] Step 104: Obtain the image to be registered and extract the key points of the image to be registered.

[0026] Step 105: Crop the image to be registered into image blocks based on the key points of the image to be registered.

[0027] Step 106: Input the image patch into the feature extraction model to obtain the feature descriptor.

[0028] Step 107: Generate the corresponding point set based on the feature descriptor.

[0029] Step 108: Determine the affine registration matrix based on the corresponding point set to obtain the registration result.

[0030] By implementing steps 100-108 above, this application can overcome the shortcomings of existing registration methods to a certain extent, obtain more discriminative feature representations, improve the registration accuracy of coronary artery images, and thus effectively solve the deficiencies of existing registration methods in terms of accuracy and efficiency.

[0031] Furthermore, based on the above description, the coronary artery affine registration method provided in this application can be applied to temporal registration of coronary artery X-ray angiography images acquired at different time points, precise localization and assessment of coronary artery stenosis areas, and real-time navigation and planning for coronary artery interventional procedures. The registration process for each application scenario is as follows: Figure 1 and Figure 2 As shown.

[0032] In one exemplary embodiment of this application, the feature extraction model provided is specifically designed for deep feature extraction tasks of coronary artery images. This feature extraction model employs a pure encoder structure, integrating multiple attention mechanisms and heterogeneous dual-path modules. In the feature space, it effectively narrows the distance between similar samples and widens the distance between dissimilar samples, thereby improving the discriminative power of feature representation and enabling the model to learn more discriminative descriptors. For example... Figure 3As shown, the feature extraction model consists of two key parts: an encoding unit and a descriptor extraction unit. As described above, the encoding unit, by stacking multiple layers of ConvSSMBlock structures and incorporating a multi-level attention mechanism, can extract feature representations of the image at different scales. Furthermore, by introducing the Leaky ReLU activation function, nonlinear modeling capabilities are enhanced. Based on this, the encoding unit includes shallow, intermediate, and deep stages. In the shallow stage, the encoding unit integrates a learnable Sobel-Aware Spatial Attention (SASA) mechanism to strengthen the perception and modeling of local details such as image edges and textures, improving the ability to analyze spatial distribution features. In the intermediate stage, a complete SDCBAM module is introduced, fusing channel and spatial dimension information. Through a collaborative enhancement mechanism, it achieves deep fusion of local structure and intermediate semantic features, improving the completeness of feature representation and contextual consistency. In the deeper stages, feature representation focuses on high-order semantic information. The encoding unit employs Dynamic Channel Gating Attention (DCGA) to effectively suppress redundant channels and enhance selective attention to task-related semantics, thereby improving the consistency and compactness of the overall semantic representation. Multi-level features are concatenated and fused through a feature fusion module to form a unified feature representation, achieving complementary advantages between different scales and semantic levels, constructing a richer and more structured feature space, and providing solid support for subsequent image registration tasks. In the descriptor extraction stage, the descriptor extraction unit integrates the fused features through convolution operations, ultimately outputting a 128-dimensional descriptor, and normalization processing ensures the stability and comparability of the descriptor in registration similarity measurement. Figure 3 In this context, Fusion represents the feature fusion module, and Conv represents the convolutional layer.

[0033] Based on the above description, the overall data processing flow of the feature extraction model provided in this application can be described as follows: The input image (or image patch) first enters the SASA mechanism and ConvSSMBlock module, which helps to highlight local details and structural information in the image at the shallow stage (i.e., low level), and strengthens the modeling of spatial distribution features such as edges and textures. Then, through a combination structure of four ConvSSMBlock modules and the SDCBAM module, spatial and channel information are taken into account at the intermediate level, and local details and intermediate semantics are enhanced synergistically. Subsequently, it is input into a combination structure of two ConvSSMBlock modules and the DCGA mechanism, which emphasizes feature channels related to the target task at the deep stage (i.e., deep level), and improves semantic consistency. Finally, average pooling layers are used to process it into multi-scale feature maps, which are then fused and normalized to form feature descriptors.

[0034] In one exemplary embodiment of this application, such as Figure 7 As shown, the ConvSSMBlock module first employs a channel partitioning strategy to divide the input features into two groups along the channel dimension. These groups are then fed into the CNN (Convolutional Neural Network) branch and the SSM branch, respectively, efficiently fusing local and global feature representations and further enhancing the network's ability to model complex structures in coronary artery images. The CNN branch (corresponding to...) Figure 7 The CNN-Branch branch focuses on local receptive field modeling, utilizing a series of small-scale convolutional operations to capture detailed textures, edge structures, and local patterns in images, enhancing the model's responsiveness to subtle changes and local anatomical structures. The SSM branch (corresponding to...) Figure 7 The SSM-Branch algorithm in this context is based on the SS2D (Selective Scanning 2D) algorithm. It employs a two-dimensional selective scanning mechanism, introduces a cross-scanning module, and uses features... Figure 4 The traversal strategy from one corner to its opposite corner effectively expands the receptive field, ensuring that each pixel in the feature map can integrate information from different directions across the entire image. This achieves efficient global information modeling without significantly increasing computational burden. The computation process of the SSM branch is as follows: Figure 8 As shown, the features extracted by the dual branches are fused along the channel dimension. A random shuffling operation breaks down the grouping restrictions between channels, effectively mitigating the information flow isolation problem that may be caused by grouped convolution, and enhancing the interactivity between channels and the overall feature expression capability. The ConvSSMBlock module, through a collaborative design of local fine-grained perception and global context fusion, can significantly improve the performance of the feature extraction model in recognizing and registering complex structural relationships in coronary artery images. Figure 7 In this context, "Split" means to group data, and "shuffle" means to shuffle the data. Figure 8 In this context, LN represents layer normalization, and Linear represents a linear layer.

[0035] In one exemplary embodiment of this application, such as Figure 4 The SDCBAM module shown integrates the SASA mechanism (such as...) Figure 5 (as shown) and DCGA mechanism (such as) Figure 6As shown in the figure, it enables efficient identification of key structures in the coronary arteries. The SDCBAM module has significant advantages in feature enhancement and selective suppression, effectively improving the discriminativeness and structural sensitivity of feature representation. Specifically, the SASA mechanism introduces a trainable Sobel operator for image gradient extraction, enabling the feature extraction model to adapt to different image distributions in edge detection tasks, thereby enhancing the modeling effect on fine-grained edge structures. The DCGA mechanism, through adaptive learning of channel weights, automatically determines the threshold for dividing high and low response channels based on the weight distribution under unsupervised conditions, effectively suppressing channels with low information contribution while retaining feature channels with significant responses, strengthening the feature extraction model's focus on key semantic features. Through coordinated regulation in both spatial and channel dimensions, the SDCBAM module can significantly improve the feature extraction model's recognition accuracy of complex coronary artery structures, providing more robust feature support for subsequent image registration. Figure 5 In Chinese, MLP stands for Multilayer Perception. Figure 6 In this context, Sobel represents the Sobel operator.

[0036] In an exemplary embodiment of this application, to further improve the accuracy of the extracted descriptors and enhance image processing performance, steps 101 and 104 above employ the SIFT algorithm to extract key points from the standardized image and the image to be registered. First, a Gaussian pyramid is constructed for the image to be registered or the standardized image, and the difference Gaussian between adjacent scales is calculated. Then, extreme points are searched in space and scale, and accurately located through interpolation. Finally, low-contrast points and edge response points are removed, returning stable key point coordinates.

[0037] In an exemplary embodiment of this application, in steps 101 and 105, the image to be registered or the standardized image can be cropped into an image block of size 64×64, centered on the key point coordinates.

[0038] In an exemplary embodiment of this application, during the process of constructing the feature extraction model in step 102, the model also needs to be trained. Therefore, experimental data needs to be acquired, and then training and testing datasets are constructed based on the acquired experimental data. In practical applications, the overall model training process may include: Step 1: Preprocess the experimental data.

[0039] 1. Image standardization: Coronary angiography images are uniformly adjusted to a resolution of 512×512 pixels to ensure consistency of input data.

[0040] 2. Noise suppression: Gaussian filtering (σ=1.5) is used to remove image noise, preserving the vascular structure while smoothing the background area.

[0041] In the formula, These represent the offset from the filter center along the horizontal or vertical direction, respectively, with σ being the standard deviation. This represents the noise suppression results.

[0042] 3. Data Normalization: This application uses a linear normalization method to scale the grayscale values ​​of the experimental dataset to the range of 0~255. The calculation formula is as follows:

[0043] In the formula, This represents the normalized pixel value. Represents the original grayscale value. and These represent the minimum and maximum grayscale values ​​of the image, respectively.

[0044] 4. Data Augmentation: The training dataset is expanded using image augmentation techniques such as random rotation, translation, and scaling to improve the generalization ability of the feature extraction model. Specific parameters are shown in Table 1. Coronary artery dataset-A, coronary artery dataset-B, and coronary artery dataset-C can be obtained through data augmentation.

[0045] Table 1 Data Augmentation Parameters

[0046] Step 2: Key point detection and image patch cropping. See the above description for the specific implementation method.

[0047] Step 3: Extract feature descriptors.

[0048] 1. Encoder Feature Extraction: Encoding units are used to handle complex image variations. The number of feature channels in the encoding unit changes from 1→16→32→32→64→64→128→128. The input to the encoding unit is a large number of image patches used for training, and the output is a depth feature map with multiple channels, the same size as the image patches, but the number of channels changes from 1 to 128.

[0049] 2. Cross-level Feature Fusion: This feature fusion module integrates feature information at different scales, effectively combining multi-level features to enhance the richness of feature representations and the semantic connections between levels. The feature fusion module consists of convolutional layers with a kernel size of 4×4 (corresponding to...). Figure 3 Conv and BN (Batch Normalization) layers (corresponding to) Figure 3 BN in the middle) and LeakyReLU layer (corresponding to Figure 3 It consists of Leaky ReLU.

[0050] 3. Feature Descriptor Generation: During the training phase, the fused feature maps are fed into a normalization layer, and then a loss function is constructed to implement feature learning. During the prediction phase, a 128-dimensional feature descriptor is generated for each keypoint. The feature descriptor contains local texture information and global structural information, and uses the L2 norm (corresponding to...) Figure 3 L2Norm in the code performs regularization on the descriptor.

[0051] In one exemplary embodiment of this application, the process of calculating the affine transformation matrix and the registration result may include: 1. Affine transformation matrix calculation. Determine the Euclidean distance between the first and second feature descriptors to obtain the similarity matrix.

[0052] 2. Obtain initial corresponding point pairs based on the similarity matrix.

[0053] 3. Use the RANSAC algorithm to filter initial corresponding point pairs to obtain matching point pairs, thus generating a corresponding point set. This process aims to remove outliers and improve matching accuracy. In practical applications, the number of iterations is set to 1000, and the threshold is 2.0 pixels.

[0054] 4. Calculate the corresponding affine transformation matrix based on the selected matching point pairs.

[0055] 5. Image deformation and resampling: Based on the calculated affine transformation matrix, bilinear interpolation is used to resample the floating image to generate an affine registered image.

[0056] 6. Affine registration quality assessment: Calculate the root mean square error (RMSE) and structural similarity index (SSIM) before and after affine registration to evaluate the affine registration effect as evaluation criteria.

[0057] In one exemplary embodiment of this application, comparative experiments are conducted to illustrate the superiority of the coronary artery affine registration method provided in this application compared to traditional registration methods. Specifically, the method provided in this application and traditional registration methods were comprehensively evaluated on three different coronary artery XA datasets. The main performance metrics included NoM, RMSE, and SSIM, and the comparison results are shown in Table 2.

[0058] Table 2 Comparison of results of the three methods on different datasets

[0059] In Table 2, the bolded data represents the best results of this indicator among different methods on the dataset. It can be seen that the method provided in this application achieves the best affine registration effect for different degrees of deformation. Compared with the traditional SIFT algorithm and the use of HardNet to extract descriptors, the method implemented in this application helps to find more effective matching points, thereby calculating the affine matrix for more accurate affine registration. The network implemented in this application achieves RMSE accuracy that is 0.1751 and 0.0605 pixels higher than SIFT and HardNet, respectively, on the coronary artery dataset-A; 0.4211 and 0.3896 pixels higher than SIFT and HardNet, respectively, on the coronary artery dataset-B; and 2.5513 and 0.6502 pixels higher than SIFT and HardNet, respectively, on the coronary artery dataset-C.

[0060] In summary, to address the technical bottlenecks of traditional registration techniques, such as insufficient accuracy, low computational efficiency, and poor adaptability to complex deformations, and to provide reliable technical support for the accurate diagnosis and treatment of coronary heart disease, this application presents a coronary artery affine registration method that constructs a multi-innovative technical system focusing on feature extraction and registration accuracy optimization. The core architecture is the EGHardNet network, which employs a pure encoder structure and constructs an efficient feature extraction path by fusing convolutional neural networks, Mamba modules, and attention mechanisms, outputting highly discriminative image descriptors. Specifically, the SDCBAM module combines learnable Sobel spatial attention and dynamically gated channel attention mechanisms to enhance the model's sensitivity and expressive ability to key anatomical structures of the coronary arteries. The ConvSSMBlock module combines the local structure modeling capabilities of CNNs with the advantages of SSM in long-range dependency information modeling, further enhancing the depth and breadth of the model's feature extraction. To fully utilize information at different semantic levels, the method also designs a cross-level feature fusion strategy, effectively integrating spatial structure information from shallow features with semantic information from deep features, thereby improving the discriminability and robustness of the descriptors during the matching stage. By organically combining the above-mentioned technical approaches, this application achieves high precision, high efficiency, and high robustness of coronary artery images in affine registration tasks, and has good clinical application prospects and promotional value.

[0061] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 9As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores coronary artery affine registration data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a coronary artery affine registration method.

[0062] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0063] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0064] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0065] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0066] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0067] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0068] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for affine registration of coronary arteries, characterized in that, include: Obtain coronary artery X-ray angiography images and perform standardization processing on the coronary artery X-ray angiography images to obtain standardized images; Key points are extracted from the standardized image, and image patches are cropped centered on the detected key points to generate training samples; An encoder structure is adopted, and a feature extraction model is constructed by fusing convolutional neural networks, Mamba modules, and attention mechanisms. The feature extraction model includes an encoding unit and a descriptor extraction unit; the encoding unit extracts feature representations of the input image at different scales by stacking multiple heterogeneous dual-path modules and enhanced attention modules, and introducing the Leaky ReLU activation function; the descriptor extraction unit includes a feature fusion module. The feature extraction model is trained using the training samples until a set condition is met, resulting in a well-trained feature extraction model. Acquire the image to be registered and extract the key points of the image to be registered; The image to be registered is cropped into image blocks based on the key points of the image to be registered; The image patch is input into the trained feature extraction model to obtain feature descriptors; Generate a corresponding point set based on the feature descriptor; Based on the corresponding point set, the affine registration matrix is ​​determined, and the registration result is obtained.

2. The coronary artery affine registration method according to claim 1, characterized in that, The heterogeneous dual-path module uses a channel partitioning strategy to divide the input features into two groups along the channel dimension: one group is input into the CNN branch and the other group is input into the SSM branch, fusing local and global features; the heterogeneous dual-path module uses downsampling operations to extract information at different scales.

3. The coronary artery affine registration method according to claim 2, characterized in that, The CNN branch is used to model local receptive fields; the SSM branch is based on the SS2D algorithm and introduces a cross-scan module on the basis of a two-dimensional selective scanning mechanism to expand the receptive field range and realize global information modeling.

4. The coronary artery affine registration method according to claim 1, characterized in that, The enhanced attention module achieves the identification of coronary artery structures by fusing the learnable Sobel spatial attention mechanism with the dynamic channel-gated attention mechanism.

5. The coronary artery affine registration method according to claim 1, characterized in that, The encoding unit includes a shallow stage, an intermediate stage, and a deep stage; In the shallow stage, a learnable Sobel spatial attention mechanism is integrated on the basis of the heterogeneous dual-path module to enhance the perception and modeling of local image information. In the intermediate layer stage, an enhanced attention module is introduced on the basis of the heterogeneous dual-path module to fuse channel and spatial dimension information and adopt a collaborative enhancement mechanism to achieve deep fusion of local structure and intermediate semantic features. In the deeper stage, a dynamic channel gating attention mechanism is adopted based on the heterogeneous dual-path module to suppress redundant channels.

6. The coronary artery affine registration method according to claim 1, characterized in that, The SIFT algorithm is used to extract key points from the standardized image and the image to be registered.

7. The coronary artery affine registration method according to claim 1, characterized in that, Generating a corresponding point set based on the feature descriptor includes: Determine the Euclidean distance between the feature descriptors to obtain the similarity matrix; Initial corresponding point pairs are obtained based on the similarity matrix; The RANSAC algorithm is used to filter the initial corresponding point pairs to obtain matching point pairs, thereby generating the corresponding point set.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the coronary affine registration method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the coronary affine registration method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the coronary affine registration method according to any one of claims 1-7.

Images