X-ray coronary image segmentation method with cross-region spatial feature consistency

Abstract

This invention discloses a cross-regional spatial feature consistency X-ray coronary artery image segmentation method, belonging to the field of image segmentation technology. The method includes: model structure design; introduction of a feature embedding module; and introduction of a pixel propagation module. This invention achieves fine segmentation of vascular structures.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and specifically to a method for segmenting X-ray coronary artery images with cross-regional spatial feature consistency. Background Technology

[0002] Currently, deep learning dominates the segmentation of blood vessels in X-ray coronary images, and supervised deep learning models have segmentation performance comparable to experts. However, such deep learning follows a fully supervised learning scheme, and the potential of the segmentation model can only be released by training the model with a large amount of annotated data, i.e., a fully supervised method. In the field of medical imaging analysis, each data acquisition point often only has a limited amount of image data and corresponding labels, which may limit the effectiveness of the learning process. In addition, supervised methods rely entirely on labels provided by experts as the only learning mode in the training process, which also brings some problems: (1) The entire raw data contains richer structural information than the labeled data, and the scarcity of data leads to the possibility of overfitting or low robustness in the training model. (2) Supervised learning can only handle a single segmentation task under the guidance of a given label. However, unlabeled data contains more useful information, such as spatial structure and related attributes, which actually provides a beneficial and important supplement to supervised learning. (3) Labeling low-contrast and complex vascular images by clinical experts is a heavy task, which may result in small datasets and high costs. Specifically, the limited number of samples in X-ray coronary image datasets makes it difficult for deep learning models to acquire sufficiently comprehensive feature information during the learning process. Due to insufficient data, spatial correlations during model training become unbalanced, potentially causing the model to overlook the overall structure of the segmented target, thus affecting its ability to accurately distinguish semantic features. Furthermore, the lack of sufficient training data can lead to slow model convergence, resulting in oversegmentation and undersegmentation. Therefore, a more flexible, efficient, and instructive method is needed to address these challenges.

[0003] While X-ray coronary artery images lack abundant labeled data, richly diverse unlabeled image data is relatively readily available. Since unlabeled image data contains far more information than sparsely labeled data, it can potentially yield more general knowledge through self-supervised learning. Self-supervised learning offers a novel paradigm that significantly narrows the gap with fully supervised methods on large computer vision benchmarks. This paradigm provides a highly productive solution for dealing with limited available annotated data. Its core idea is to pre-train using unsupervised strategies to obtain more useful and discriminative representations of the data samples. Subsequently, the pre-trained model is fine-tuned using limited labeled samples to make it suitable for practical tasks such as segmentation and classification. This method provides a flexible and efficient approach to handling data scarcity, offering a novel perspective on addressing challenges. The contrastive learning method within this approach is based on the principle of similarity maximization, requiring efficient generation of positive samples (relevant images) and negative samples (unrelated images) for pre-training. While these strategies have achieved significant success in the natural image domain, they have shown certain limitations in X-ray coronary artery image analysis.

[0004] Contrastive learning models rely on memorizing deep-level features for prediction, while X-ray coronary artery fine segmentation is essentially a fine-grained image refinement task, where discriminative features often focus on subtle structural regions and boundaries within the image. Therefore, more detailed shallow features are needed to provide an accurate visual description of these structures. However, this focus on detailed information in effective feature acquisition can lead to longer training times, slower convergence, and a greater susceptibility to local minima and collapsing solutions. Summary of the Invention

[0005] The main objective of this invention is to provide a cross-regional spatial feature consistency X-ray coronary artery image segmentation method to solve the problem of insufficient number of labeled samples in X-ray coronary artery images.

[0006] The technical solution adopted in this invention is: a cross-regional spatial feature consistency X-ray coronary artery image segmentation method, comprising: Model structure design; Introducing a feature embedding module; A pixel propagation module is introduced.

[0007] Furthermore, the model structure design includes: A cross-region feature consistency network (FCCN) is constructed. The FCCN structure includes two main parts: contrastive learning pre-training and fine-grained segmentation. Two randomly perturbed views from the same image are input into the backbone network UNet for feature encoding; In the process of contrastive learning, MLP is introduced as a projection network to form the structure of contrastive learning; Projection networks utilize MLPs to generate multiple feature vectors through fully connected layers; The output of the feature embedding module is first mapped into a feature vector through global average pooling, and then the representation is extracted through an FC layer, batch normalization, and the ReLU activation function. MLP not only generates regression targets for training the pixel propagation module, but also updates itself by receiving only dynamic weights from the pixel propagation module; By comparing the output representation vectors, the model can identify different transformation features of the same image; this process, through a contrastive learning mechanism, enables the model to learn more generalized feature representations. For each image Two sets of randomly combined perturbation operators are used. The generated enhanced view is as follows: and To update parameters in the online network and perform exponential moving average (EMA) updates in the target network; After passing the attention module and aligning the representation embeddings, a global position loss is used in conjunction. and global channel loss To minimize the consistency between features; in the final linear inference stage, only the encoder is retained and effective segmentation is performed.

[0008] Furthermore, the feature embedding module includes: The feature embedding module is introduced to focus on feature space invariance; in the feature embedding module, the feature map Mapped to , , Three mapping feature maps, , , Represents the weight matrix; mapping feature map , Information exchange, through matrix multiplication and The activation function is used to compute spatial self-attention weights to generate a weighted feature map; this is then compared with the feature map. Reshaping Form Multiplication yields spatial attention interaction features; parallel additive computation is employed, i.e., a skip connection is added to the original self-attention path; this module is designed to highlight both local and global feature representations of the feature map by adding attention encoding to the original feature map. (12) (13).

[0009] Furthermore, the introduced pixel propagation module includes: The pixel propagation module integrates the self-attention mechanism and multi-head attention of Transformer to establish strong connections between different regions of the input image; Considering input , Represents the number of pixels. This represents the feature dimension; for the input Perform three linear transformations respectively to obtain the query results. , key Sum : (14) in, , , It is the weight matrix learned, used to guide the input features to be mapped to a new representation space; In the new representation space, features are aggregated and refined by channel. Layer normalization operation was performed. It measures the correlation between query features and key features; Introducing a correlation measurement block To calculate the affinity matrix and measure the distinct responses between the query location and features across all locations; to calculate the attention distribution of the target pixel. Then through The scaling dot product attention operation allows the model to incorporate positional tradeoffs into the feature representation; and introduces... ; (15) (16) (17) (18).

[0010] Furthermore, it also includes the design of loss functions, specifically: contrastive learning pre-training loss and segmentation loss; A Gaussian kernel is introduced for vascular feature extraction; Assuming the projection output is Pixel propagation output is Now, communication between pixels is established based on pixel similarity, and the similarity of positional features is represented by a Gaussian kernel as follows: (19) Meanwhile, by using the L2 norm to measure the correlation between two samples, the network is encouraged to achieve high similarity values ​​for similar samples, thereby learning meaningful feature representations. (20) Combining location awareness and similarity awareness, we obtain the total loss function for contrastive learning pre-training: (twenty one).

[0011] Advantages of this invention: This invention constructs a Cross-Regional Feature Consistency Network (FCCN). First, based on an asymmetric contrastive learning framework, data perturbation is used to increase the diversity and quantity of samples. UNet is employed as the backbone of the contrastive learning framework, utilizing a Feature Embedding Module (FEM) to focus on consistency in the feature space and extract more discriminative feature representations. A Pixel Propagation Module (PPM) is introduced to replace the MLP in the predictor, achieving pixel propagation across regions. To further improve the network's sensitivity to vascular structures, a contrastive learning loss is designed, combining location awareness and similarity awareness. This allows the model to pay more attention to subtle changes in vascular structures during inference and backpropagation, thereby achieving fine segmentation of vascular structures.

[0012] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0013] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is a comparison framework diagram of BYOL; Figure 2 This is the cross-regional feature consistency network of the present invention; Figure 3 This is the feature embedding module of the present invention; Figure 4 This is the pixel propagation module of the present invention; Figure 5 This is a visualization of the class activation graph of the FEM module of the present invention (wherein, (a) the original image, (b) the CAM graph without FEM, and (c) the CAM graph with FEM). Figure 6 This is an experimental diagram of the position-aware loss function ablation method of the present invention; Figure 7 These are the PR curves on different test sets of this invention; Figure 8This is a performance comparison result of the present invention under different training set ratios; Figure 9 These are different methods of the present invention compared visually on the DAG1 test set; Figure 10 The different methods of the present invention are compared visually on the test set JMA2 (wherein, (a) the original image, (b) the ground truth image, (c) CIDN, (d) SwAV, (e) SimSiam, (f) NNCLR, (g) BYOL, (h) FCCN). Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Components of a contrastive learning model Comparative learning frameworks: Contrastive learning, as a widely applied technique in self-supervised learning models, has been extensively used in tasks such as anomaly detection, object tracking, and depth estimation. In recent years, a series of concise and efficient contrastive learning frameworks have emerged, achieving significant performance improvements for unsupervised models in visual tasks. The main idea behind these frameworks is to guide the network model to shorten the representation distance between positive sample pairs while minimizing the interference of negative samples. For example, a simple visual representation contrastive learning framework called SimCLR was proposed. It treats two augmented views of the same image as positive sample pairs, mixes them with negative samples, and trains them discriminatively. SimCLR elevates the performance of self-supervised models to a new level. Based on SimCLR, the Momentum Contrastive Learning (MoCo) framework was proposed, significantly reducing memory resource overhead, and MoCov3 was derived through iterative improvements. Other researchers have replaced the computationally intensive sample comparison process with a comparison of clustering results under different augmented views, further improving clustering-based deep self-supervised learning methods (SwAV). Minimizing the differences between positive sample pairs can promote modeling invariance, and further separation of negative sample pairs is considered a necessary condition to prevent the representation from collapsing.

[0017] However, the aforementioned framework not only increases computational costs due to its requirement for a large number of positive and negative sample pairs, but also presents challenges in acquiring X-ray coronary image samples. BYOL (Bootstrap Your Own Latent), on the other hand, is a self-supervised learning method that does not require negative samples. Its core idea is to maximize the mutual information between different views of the same data, encouraging the model to learn useful and invariant features. The consistency constraint between the two views ensures that the model representation remains consistent across views without requiring a large number of negative sample comparisons or memory queues. Its model aims to make the representations from two views of the same image as similar as possible, thereby learning useful features; however, it needs a mechanism to avoid collapse. Therefore, BYOL uses a prediction network that breaks the symmetry between the two networks and forces an exponential moving average to be applied to the weights of the target network to slow down the update speed of the weights on the target network. BYOL aims to improve the quality of feature representations without requiring actual negative sample comparisons through data augmentation, contrastive loss, and training with two views. The advantage of this method is that it does not rely on explicit label information, making it particularly suitable for self-supervised learning tasks. Therefore, the goal of self-supervised learning (BYOL) is to learn useful feature representations by comparing information between different views so that these features can be effectively utilized in subsequent supervised tasks.

[0018] Representation learning in BYOL models: In self-supervised learning methods such as contrastive learning, especially when training data is insufficiently diverse or the training process is unstable, representation collapse is a significant problem. This phenomenon can cause the model to map different inputs to similar or identical representations during training, thus reducing the model's ability to distinguish input features. In contrastive learning, the lack of negative samples can lead to representation collapse. The presence of negative samples helps the model learn to differentiate between different samples, resulting in rich feature representations. Without negative samples, the model may use simplistic representations instead of truly learning useful features. Therefore, ensuring sufficient diversity and richness of training data is crucial to preventing representation collapse.

[0019] BYOL model, such as Figure 1 As shown, the challenge of representation collapse is cleverly addressed by introducing a combination of a target network and an online predictor network. The target network is responsible for providing stable representations, while the online predictor guides the representation learning process by maximizing their mutual information. This design allows the model to maintain the diversity and richness of representations during learning, avoiding premature convergence to simplistic representations. Furthermore, by stopping gradient backpropagation of the online predictor, the model effectively decouples the target network and the predictor network, further improving performance.

[0020] Assuming the backbone network and projection network are approximated as linear mappings, the loss minimized by BYOL and the linear predictor can be expressed as: (1) Here, Indicates the input image The expected value of the distribution, , They are The output of the two enhanced views after being processed by the projection network. The weights represent the linear predictor. express The gradient did not propagate back through the second path. If each They are all standardized, that is ,So It can be replaced with a small batch The average value, and defined as follows: , .

[0021] At this time there is .

[0022] Next, we will elaborate on (1): (2) The loss function applies to the predictor weights respectively and online projection network Differentiate: (3) (4) Treating the changing network weights during training as a time variable, i.e. , The network adjusts its weights during training. To minimize the loss using decay and gradient descent, then: (5) (6) Substitute (5) and (6) into the following equation: (7) Substituting (3) and (4) into (7) and rearranging, we get: (8) For equation (8), Integrating, we get: (9) in It is the initial constant matrix; multiply both sides by . get: (10) when When a certain value is reached, (10) can be simplified to: (11) This means that minimizing the loss function makes Approaching , Non-central covariance .

[0023] In summary, weight decay promotes a balance between the predictor and the online network. Even without negative samples during training, the BYOL model, by minimizing the loss function, still allows the projected features to approximate the non-center covariance of positive samples. This means the model can effectively learn to remove correlations between data points, resulting in more discriminative feature representations, even without the intervention of negative samples. In this case, the model is trained based on the relationships between positive samples, achieving decorrelation, and the absence of negative samples does not affect the achievement of this goal.

[0024] Inspired by this, an asymmetric contrastive learning network is used to construct a contrastive learning framework. By using an online predictor and stopping the gradient of the target network, the correlation between negative sample pairs is weakened, enabling the model to learn more general and abstract feature representations. In the X-ray coronary artery image segmentation task, strategies such as improving the asymmetric contrastive learning network model structure will be employed to avoid or mitigate the impact of representation collapse, ensuring that the model can learn more discriminative feature representations.

[0025] Data augmentation: In contrastive learning, data augmentation, by increasing the diversity of input data, helps the model focus on extracting global representations from the data, improving the generality of the learned representations, and thus learning more generalized feature representations. A series of data augmentation techniques, such as rotation, flipping, scaling, and cropping, were implemented on the input image. To enhance the model's adaptability to different viewing angles, random rotation and flipping operations were introduced. The random rotation range was set between 0 and 360 degrees, while the flipping operation included both horizontal and vertical directions. Furthermore, the image was randomly scaled and cropped to simulate observation conditions at different distances and with varying field of view. The scaling ratio and cropping size were randomly selected according to a predefined probability distribution. The purpose of these strategies is to enable the proposed model to learn stable feature representations under various image transformations, thereby improving the model's adaptability to input perturbations.

[0026] Cross-regional spatial feature consistency methods: Self-supervised representation learning of X-ray coronary artery images is achieved through a cross-regional spatial feature consistency method. The framework's construction is explained in three aspects: First, the network structure of the proposed model is introduced in detail. Second, the key roles of the feature embedding module and pixel propagation module in the model are discussed. Finally, the loss function used is explained in detail.

[0027] Model structure design: The asymmetric contrastive learning framework based on positive samples encompasses a dual learning path for the same set of input images, comparing the feature representations extracted through these two paths. In this process, the network gradually narrows the gap between these representations, enabling it to learn the unique discriminative features of each image. As... Figure 2 The proposed method constructs a cross-region feature consistency network (FCCN), which consists of two main parts: contrastive learning pre-training and fine-grained segmentation.

[0028] First, features are encoded by inputting two randomly perturbed views from the same image into the backbone network UNet. A feature embedding module focuses the feature space, emphasizing both local and global feature representations. In the contrastive learning process, an MLP is introduced as a projection network, forming the contrastive learning structure. The projection network utilizes the MLP to generate multiple feature vectors through fully connected layers. Specifically, the output of the feature embedding module is first mapped to a feature vector through global average pooling (GAP), followed by representation extraction through fully connected layers, batch normalization (BN), and the ReLU activation function. The MLP not only generates regression targets for training the pixel propagation module but also updates itself by receiving dynamic weights only from the pixel propagation module. Finally, by comparing the output representation vectors, the model can recognize different transformation features of the same image, such as... Figure 2 As shown. This process, through a contrastive learning mechanism, enables the model to learn more generalized feature representations, thereby achieving better performance when processing different transformations of the same image.

[0029] For each image Using two sets of randomly combined perturbation operators The generated enhanced view is as follows: and The parameters are updated in the online network, and an exponential moving average (EMA) update is performed in the target network. After passing the attention module and aligning the representation embedding, a global position loss is jointly used. and global channel loss This minimizes the consistency between features. In the final linear inference stage, only the encoder is retained for effective segmentation.

[0030] Feature embedding module: To preserve the original encoded information of the feature maps, better maintain the geometric shape of the features, and enhance the understanding of image structure, a Feature Embedding Module (FEM) is introduced to focus on feature space invariance. For example... Figure 3 In the feature embedding module, the feature map Mapped to , , Three mapping feature maps, , , This represents the weight matrix. Mapping the feature map... , Information exchange, through matrix multiplication and An activation function is used to compute spatial self-attention weights, thereby generating a weighted feature map. This map is then compared with the feature map. Reshaping Form Multiplication yields spatial attention interaction features. Finally, parallel additive computation is employed, adding a skip connection to the original self-attention path. This module's design highlights both local and global feature representations of the feature map by adding attention encoding to the original feature map.

[0031] (12) (13) Pixel propagation module: The Pixel Propagation Module (PPM) integrates the self-attention mechanism and multi-head attention of the Transformer, cleverly establishing strong connections between different regions of the input image. The PPM can establish associations between feature representations to effectively capture contextual information in the image. Its fundamental goal is to fully leverage the powerful features of the Transformer, ensuring that each pixel benefits from global contextual information through the ingenious feature propagation strategy. Figure 4 As shown.

[0032] Considering input , Represents the number of pixels. This represents the feature dimension. First, for the input... Perform three linear transformations respectively to obtain the query results. , key Sum : (14) in, , , It is the weight matrix learned, used to guide the input features to be mapped to a new representation space; In the new representation space, features are aggregated and refined by channel. Layer normalization operation was performed. This measures the correlation between query features and key features. To achieve effective constraints on target pixels, a correlation measurement block is introduced. This is used to calculate the affinity matrix and measure the distinct responses between the query location and features across all locations. The attention distribution for the target pixels is also calculated. Then through The scaled dot product attention operation allows the model to incorporate positional tradeoffs into the feature representation. It also introduces... This allows the model to better adapt to complex data distributions and feature relationships.

[0033] (15) (16) (17) (18) Design of loss function Due to the imbalance in X-ray coronary image data, with significant differences in the proportion of blood vessels to non-blood vessels, improving the model's loss function can enhance its ability to learn the target and thus improve its adaptability to imbalanced data distributions. The model uses two loss functions: contrastive learning pre-training loss and segmentation loss. The segmentation loss follows the established loss function, and the design of the contrastive learning pre-training loss function will be described. Since similarity calculation is crucial for loss function calculation in contrastive learning, the original loss calculation is extended by defining the target using the proposed joint minimization loss.

[0034] For X-ray coronary images, blood vessels typically exhibit distinct features within local regions of the image. The shape and structure of blood vessels often change within these local areas, and vessel boundaries and textures provide valuable information for later diagnosis. Considering local relationships, enhanced texture information, and spatial location, introducing a Gaussian kernel for blood vessel feature extraction can improve sensitivity to vascular structure. The Gaussian kernel focuses more on pixels within local regions when calculating similarity, rather than simply considering all pixels globally. Assuming the projection output is... Pixel propagation output is Now, communication between pixels is established based on pixel similarity, and the similarity of positional features is represented by a Gaussian kernel as follows: (19) Meanwhile, by using the L2 norm to measure the correlation between two samples, the network is encouraged to achieve high similarity values ​​for similar samples, thereby learning meaningful feature representations.

[0035] (20) Combining location awareness and similarity awareness, we obtain the total loss function for contrastive learning pre-training: (twenty one) Experimental Results and Analysis: First, the dataset used in the experiment is introduced, and the experimental environment is described in detail. Next, in-depth qualitative and quantitative analysis experiments are conducted on the two key modules of the proposed X-ray coronary artery image segmentation model, FEM and PPM, to ensure their effectiveness is verified. Finally, the segmentation performance of the model is evaluated on various datasets, and a comprehensive comparison with current leading models is presented, yielding a detailed comparative report.

[0036] Dataset and Experiment Setup: All experiments used the same two datasets, JMA2 and DCA1, with all datasets divided into training and testing sets in the same proportion: 80% for training and 20% for testing. After training, the model's performance was tested by freezing the weights of the backbone network. Furthermore, state-of-the-art contrastive learning models were reproduced using code provided by the authors in the literature and resource repositories, employing the same backbone network to ensure a fair comparison of classification performance.

[0037] To standardize the experiments, cubic interpolation was used to adjust the image resolution of all datasets to [value missing]. The experimental environment was deployed on a server running Ubuntu 18.04.5. All experiments were run using Python 3.8, and the GPU was an NVIDIA Quadro RTX4000. The backbone network was uniformly UNet, and the SGD optimizer was used to learn the optimal parameters, with the SGD momentum set to 0.9. The initial learning rate was... The network is trained according to the cosine decay period, and its weights are frozen after training is completed.

[0038] In the performance comparison experiments, various training and test set ratios and cross-validation schemes were employed to comprehensively evaluate the performance of the proposed model. To simulate the case of limited labeled data, the study considered 30%, 50%, and 80% cross-validation of the training dataset to fine-tune the segmentation model and obtain better segmentation results. An early stopping strategy was also used during the training phase to avoid overfitting; the training process was terminated when the loss function no longer decreased.

[0039] In addition, a series of performance evaluation metrics were used to test the generalization ability of the proposed model, including precision, recall, specificity, F1 score, and accuracy.

[0040] Qualitative and quantitative experimental analysis: Effectiveness analysis of feature embedding module To assist in verifying the effectiveness of the Feature Embedding Module (FEM), this section uses the GradCAM method for visualization analysis. The visualization results of the FEM module are observed by randomly selecting three images, such as... Figure 5 As shown in the results, the effective FEM module produces strong activations near vascular structures. Compared to images without FEM processing, the image samples focus less on vascular artifacts and more on the vascular region. Furthermore, the FEM module prioritizes feature consistency in the image. The highlighted areas of GradCAM highly match the actual locations of blood vessels, further validating the visual effectiveness of the FEM method compared to images without FEM processing. This visualization analysis provides a more intuitive understanding of the role of the FEM module in vessel segmentation tasks, offering direct evidence for improving model performance.

[0041] Effectiveness analysis of pixel propagation module To verify the effectiveness of the Pixel Propagation Module (PPM), this section presents an ablation experiment. The PPM is used to replace the predictor MLP in the traditional BYOL contrast method, and the performance of the two settings is compared, focusing on metrics such as F1 score, sensitivity (Sen), specificity (Spe), and accuracy (Acc).

[0042] In the ablation experiments (Table 1), the model using PPM as the predictor showed superior performance in terms of F1 score, sensitivity, specificity, and accuracy. The introduction of PPM successfully improved the model's performance on the vessel segmentation task. Specifically, the improved F1 score indicates a balance achieved by the model while simultaneously considering precision and recall; higher sensitivity and specificity indicate more accurate identification of positive samples, while higher accuracy further confirms the overall performance improvement.

[0043] Table 1 PPM Ablation Experiment

[0044] Effectiveness analysis of the location-aware loss function This section analyzes the loss function of segmentation results using random samples. Four images were randomly selected as examples. Figure 6The results were compared between the actual values ​​and the model predictions to observe the boundary fitting effect. The first row does not use the position-aware loss function, while the second row does. In the segmentation results, green represents the boundary of the actual values, and red represents the boundary of the model's predictions. By comparing the two rows, it is clear that the second row's predictions have a higher degree of overlap with the actual boundaries in some areas, indicating a better fit, as marked in yellow in the figure. This shows that the proposed position-aware loss function plays a positive role in improving the attention to boundary position details. Through position-aware loss, local features of blood vessels can be better captured, texture information enhanced, and the importance of spatial location maintained, thereby improving the sensitivity to vascular structures and maximizing feature consistency.

[0045] Performance Comparison Quantitative analysis of network model performance This section compares the quantitative evaluation with other representative segmentation methods, including SimSiam, SwAv, NNCLR, and BYOL. All of these models were retrained on the JMA2 dataset in the same environment to obtain the corresponding results.

[0046] Table 2 Comparison of performance metrics of different models on the DCA1 test set

[0047] Quantitative analysis, as shown in Tables 2-3, revealed that the FCCN model achieved significant improvements across multiple performance metrics. Specifically, the FCCN model exhibited superior performance in F1 score, sensitivity (Sen), specificity (Spe), and accuracy (Acc). In particular, the linear fine-tuning results of the FCCN model achieved performance comparable to supervised learning. This indicates that the proposed method possesses higher accuracy and better overall performance in vessel segmentation tasks. Compared to the contrasting models, the FCCN model is more suitable for processing X-ray coronary artery images, more accurately capturing vascular structures and extracting information about abnormal details. This quantitative analysis further validates the effectiveness and superiority of the proposed method.

[0048] Table 3 Comparison of performance metrics of different models on the JMA2 test set

[0049] exist Figure 7In the PR curve, the horizontal axis represents recall, and the vertical axis represents precision. In comparisons with other pre-trained models, the PR curve of this model is more convex, indicating a better balance between different recall and precision levels. Especially at high recall, the FCCN model performs better than other models, capturing vascular structures more comprehensively. Furthermore, the FCCN model maintains high precision within a certain recall range, demonstrating high efficiency with a low false positive rate. Overall, the FCCN model exhibits better performance than other models in X-ray coronary artery image segmentation, providing strong support for improving the accuracy of automatic segmentation.

[0050] To improve the segmentation performance of FCCN, we used 30%, 50%, 80%, and 100% of the data volume to verify the effectiveness of comparative learning of FCCN. Figure 8 The chart clearly demonstrates the performance variation of FCCN under different training set proportions. As the percentage of training samples increases, the model's performance, such as the F1 score, shows a steady upward trend, reaching optimal performance with 100% training set. Notably, the performance improvement is more significant with relatively small training data (30% and 50%), highlighting the superior performance of contrastive learning in reducing label dependence. This means the model attempts to learn useful features by comparing unlabeled data with its own variants. The proposed method achieves excellent performance in segmentation with only a small amount of labeled data, providing significant assistance in situations where X-ray coronary artery image annotation is difficult. In summary, this chart shows that contrastive learning can achieve significant performance improvements even with relatively small training data. Furthermore, the performance improvement continues as the training set proportion increases, providing room for further optimization of model performance.

[0051] Network Model Complexity Analysis To gain a deeper understanding of the performance and resource consumption of the FCCN backbone network and segmentation network in applications, Table 4 shows a comparative experiment on the complexity of the two backbone networks and segmentation networks of the FCCN model, focusing mainly on key aspects such as the number of parameters and computational complexity.

[0052] Table 4. Complexity Analysis of the FCCN Model

[0053] Based on the results in Tables 2, 3, and 4, it is found that although UNet++ offers more parameters and computational complexity compared to UNet, it does not provide a significant performance advantage for X-ray coronary artery image segmentation models. Conversely, UNet, with its fewer parameters and lower computational complexity, is more suitable for the deployment of lightweight and efficient models in X-ray coronary artery image segmentation. Therefore, UNet is selected as the backbone network for X-ray coronary artery image segmentation to achieve optimal performance.

[0054] Qualitative analysis of network model performance In the qualitative analysis of network model performance, the focus is on image quality and the accuracy of segmentation results. Reconstructed images are visualized to verify whether different contrastive pre-trained models can successfully learn the latent representation of the input data. This section selects a series of representative images for qualitative analysis, such as... Figures 9-10 As shown, by intuitively observing the images generated by each preprocessing model, the FCCN model excels in enhancing the details of vascular structures, making the vascular structures in the images clearer and more discernible. This indicates that the features obtained through self-supervised learning help reveal the details of vascular structures. Observing the boundary between real and predicted vessels, it is found that the FCCN model is highly sensitive to vessel recognition. Compared to other contrastive learning models, FCCN performs well in providing sensitive visual cues for correct boundaries. Visualization results show that the FCCN model can help the network focus on foreground regions, better fit vessel boundaries, and ensure more accurate segmentation results. In particular, from Figure 9 The third image shows that the FCCN shows the highest focus on the actual vascular structure; while from Figure 10 As can be seen from the fourth image, FCCN is most effective at suppressing interfering factors. Overall, FCCN performs comparably to the supervised learning model CIDN.

[0055] The success of supervised deep learning relies on a large number of labeled samples. To address the problem of insufficient labeled samples in X-ray coronary artery images, a cross-regional spatial feature consistency method is proposed, and a Cross-Regional Feature Consistency Network (FCCN) is constructed. First, an asymmetric contrastive learning framework is used as the foundation, increasing the diversity and quantity of samples through data perturbation. UNet is adopted as the backbone of the contrastive learning framework, utilizing a Feature Embedding Module (FEM) to focus on feature space consistency and extract more discriminative feature representations. A Pixel Propagation Module (PPM) is introduced to replace the MLP in the predictor, achieving pixel propagation across regions. To further improve the network's sensitivity to vascular structures, a contrastive learning loss is designed by combining location awareness and similarity awareness, enabling the model to pay more attention to subtle changes in vascular structures during inference and backpropagation, thereby achieving fine segmentation of vascular structures.

[0056] This invention constructs a Feature-Consistent Cross-Region Network (FCCN), which uses asymmetric contrastive learning as its basic framework and UNet as its backbone. Within FCCN, a Feature Embedding Module (FEM) is designed. This module focuses on the consistency of the feature space, enhancing the geometric structure of features through specific embedding strategies, enabling the network to more accurately capture and represent key structural information in images. A Pixel Propagation Module (PPM) is designed to establish cross-regional associations in the image, facilitating the understanding and utilization of global contextual information. A contrastive learning loss combining location awareness and similarity awareness is designed, improving FCCN's sensitivity to vascular structures during inference and backpropagation. Through appropriate data perturbation, FCCN can naturally construct pixel-level positive samples without involving large-scale negative samples. Implicit comparisons are performed through gradient stopping operations, avoiding explicit negative sample selection, simplifying the training process, and allowing the network to focus more on the fine segmentation of vascular structures.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for segmenting X-ray coronary artery images with cross-regional spatial feature consistency, characterized in that, include: Model structure design; Introducing a feature embedding module; A pixel propagation module is introduced.

2. The X-ray coronary artery image segmentation method with cross-regional spatial feature consistency according to claim 1, characterized in that, The model structure design includes: A cross-region feature consistency network (FCCN) is constructed. The FCCN structure includes two main parts: contrastive learning pre-training and fine-grained segmentation. Two randomly perturbed views from the same image are input into the backbone network UNet for feature encoding; In the process of contrastive learning, MLP is introduced as a projection network to form the structure of contrastive learning; Projection networks utilize MLPs to generate multiple feature vectors through fully connected layers; The output of the feature embedding module is first mapped into a feature vector through global average pooling, and then the representation is extracted through an FC layer, batch normalization, and the ReLU activation function. MLP not only generates regression targets for training the pixel propagation module, but also updates itself by receiving only dynamic weights from the pixel propagation module; By comparing the output representation vectors, the model can identify different transformation features of the same image; this process, through a contrastive learning mechanism, enables the model to learn more generalized feature representations. For each image Two sets of randomly combined perturbation operators are used. The generated enhanced view is as follows: and To update parameters in the online network and perform exponential moving average (EMA) updates in the target network; After passing the attention module and aligning the representation embeddings, a global position loss is used in conjunction. and global channel loss To minimize the consistency between features; in the final linear inference stage, only the encoder is retained and effective segmentation is performed.

3. The X-ray coronary artery image segmentation method with cross-regional spatial feature consistency according to claim 1, characterized in that, The introduced feature embedding module includes: The feature embedding module is introduced to focus on feature space invariance; in the feature embedding module, the feature map Mapped to , , Three mapping feature maps, , , Represents the weight matrix; mapping feature map , Information exchange, through matrix multiplication and The activation function is used to compute spatial self-attention weights to generate a weighted feature map; this is then compared with the feature map. Reshaping Form Multiplication yields spatial attention interaction features; parallel additive computation is employed, i.e., a skip connection is added to the original self-attention path; this module is designed to highlight both local and global feature representations of the feature map by adding attention encoding to the original feature map. (12) (13)。 4. The X-ray coronary artery image segmentation method with cross-regional spatial feature consistency according to claim 1, characterized in that, The introduced pixel propagation module includes: The pixel propagation module integrates the self-attention mechanism and multi-head attention of Transformer to establish strong connections between different regions of the input image; Considering input , Represents the number of pixels. This represents the feature dimension; for the input Perform three linear transformations respectively to obtain the query results. , key Sum : (14) in, , , It is the weight matrix learned, used to guide the input features to be mapped to a new representation space; In the new representation space, features are aggregated and refined by channel. Layer normalization operation was performed. It measures the correlation between query features and key features; Introducing a correlation measurement block To calculate the affinity matrix and measure the distinct responses between the query location and features across all locations; to calculate the attention distribution of the target pixel. Then through The scaling dot product attention operation allows the model to incorporate positional tradeoffs into the feature representation; and introduces... ; (15) (16) (17) (18)。 5. The X-ray coronary artery image segmentation method with cross-regional spatial feature consistency according to claim 1, characterized in that, It also includes loss function design, specifically: contrastive learning pre-training loss and segmentation loss; A Gaussian kernel is introduced for vascular feature extraction; Assuming the projection output is Pixel propagation output is Now, communication between pixels is established based on pixel similarity, and the similarity of positional features is represented by a Gaussian kernel as follows: (19) Meanwhile, by using the L2 norm to measure the correlation between two samples, the network is encouraged to achieve high similarity values ​​for similar samples, thereby learning meaningful feature representations. (20) Combining location awareness and similarity awareness, we obtain the total loss function for contrastive learning pre-training: (21)。

Images