A model training method, an X-ray blood vessel image efficient segmentation method and a medium

Abstract

The application relates to a model training method, an X-ray blood vessel image efficient segmentation method and a medium, and belongs to the technical field of computer-aided diagnosis and medical image processing. In order to solve the segmentation problem of the X-ray blood vessel image, the application provides a model training method, which comprises the following steps: acquiring an X-ray blood vessel sample image and performing combined data enhancement preprocessing; inputting the preprocessed sample image into an X-ray blood vessel image efficient segmentation network model based on frequency domain information guidance and topological relation enhancement, and outputting a blood vessel segmentation prediction result; adopting a standard binary cross-entropy loss function as a segmentation prediction network target optimization function, and calculating the loss value between the blood vessel segmentation prediction result and a preprocessed sample image label; based on the loss value, the parameters of the X-ray blood vessel image efficient segmentation network model based on the frequency domain information guidance and the topological relation enhancement are optimized and trained to obtain a target segmentation model. The image segmentation accuracy and robustness can be improved.

Description

Technical Field

[0001] This invention relates to a model training method, an efficient segmentation method for X-ray blood vessel images, and a medium, belonging to the fields of computer-aided diagnosis and medical image processing technology. Background Technology

[0002] Hypertension is the leading risk factor for death worldwide. In my country, the number of people with hypertension is nearly 300 million, but the control rate is less than 20%. Primary aldosteronism is a common cause of secondary hypertension, with an incidence rate as high as 20% in patients with refractory hypertension. Selective Adrenal Artery Embolization (SAAE), as a minimally invasive interventional treatment, has the advantages of minimal trauma and rapid recovery, providing a feasible treatment option for patients with PA (Pulmonary Arterial Hypertension) who are not suitable for surgery.

[0003] Postoperative imaging changes are a key indicator for assessing the effectiveness of SAAE treatment, but traditional assessment methods rely on physician experience and observation, which are highly subjective and lack accuracy. Deep learning-based image segmentation technology can quantify vascular changes before and after surgery, providing an objective basis for efficacy evaluation. However, X-ray vascular images suffer from low contrast, complex backgrounds, numerous noise artifacts, and delicate and complex vascular structures, posing significant challenges to the segmentation task.

[0004] Early blood vessel segmentation methods (such as thresholding and region growing) require high image quality, necessitate manual intervention, and exhibit poor generalization ability. While existing deep learning-based segmentation algorithms (such as U-Net and FR-UNet) offer improvements, they still have shortcomings: relying solely on spatial feature extraction makes it difficult to effectively distinguish low-contrast blood vessels from the background; neglecting the topological connectivity of blood vessels easily leads to the loss or breakage of small branches; and insufficient robustness to noise and artifacts affects segmentation accuracy. Therefore, there is an urgent need for an efficient segmentation method that balances feature representation capability with topological structure preservation to solve the segmentation challenges of X-ray blood vessel images. Summary of the Invention

[0005] To overcome the shortcomings of existing segmentation methods that rely solely on spatial feature extraction, making it difficult to effectively distinguish low-contrast blood vessels from the background and resulting in limited accuracy and robustness, this application provides a model training method, an efficient segmentation method for X-ray blood vessel images, and a medium. Through the synergistic effect of frequency domain information guidance and topological relationship enhancement, it effectively solves the segmentation challenges of low contrast, complex backgrounds, and high noise levels in X-ray blood vessel images, improving segmentation accuracy and robustness, and providing a scientific quantitative basis for clinical diagnosis and treatment.

[0006] An embodiment of the first aspect of this application provides a model training method, including the following steps:

[0007] X-ray blood vessel sample images were acquired and combined data augmentation preprocessing was performed to obtain preprocessed sample images;

[0008] The preprocessed sample images are input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology enhancement, and the blood vessel segmentation prediction results are output.

[0009] The standard binary cross-entropy loss function is used as the objective optimization function of the segmentation prediction network to calculate the loss value between the blood vessel segmentation prediction result and the preprocessed sample image label.

[0010] Based on the loss value, the parameters of the efficient X-ray blood vessel image segmentation network model guided by frequency domain information and enhanced by topological relationship are optimized and trained to obtain the target segmentation model.

[0011] In some embodiments, the efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology enhancement includes a feature extraction module, a feature fusion module, a decoding module, and a topology enhancement and segmentation module. The feature extraction module extracts spatial features and frequency domain edge features of the preprocessed sample image in parallel. The feature fusion module performs adaptive weighted fusion of the extracted spatial features and frequency domain edge features. The decoding module combines upsampling and skip connections to restore the resolution of the X-ray blood vessel image. The topology enhancement and segmentation module enhances the topological connectivity of blood vessels through spatial-channel graph convolution and outputs the blood vessel segmentation prediction result.

[0012] In some embodiments, the feature extraction module includes five spatial encoders, one convolutional encoder, and four frequency domain encoders. The five spatial encoders are connected in sequence. The first spatial encoder obtains a first spatial feature with 64 output channels through a convolutional layer and a multi-scale feature extraction layer. The second to fifth spatial encoders each obtain a second, third, fourth, and fifth spatial feature with 128, 256, 512, and 512 output channels, respectively, through a convolutional layer, a downsampling layer, and a multi-scale feature extraction layer. The convolutional encoder and the four frequency domain encoders are connected in sequence. The convolutional encoder consists of two 3×3 convolutions and one 1×1 convolution stacked together, outputting convolutional features. The four frequency domain encoders each include a wavelet transform layer and an edge extraction layer, outputting a first, second, third, and fourth frequency domain feature, respectively, in sequence.

[0013] In some embodiments, in the spatial encoder, the downsampling layer adopts a 3×3 convolution and max pooling operation structure, and the multi-scale feature extraction layer adopts a parallel structure of 3×3 regular convolution, 5×5 convolution and dilated convolution with dilation coefficients of 2 and 4.

[0014] In some embodiments, in the frequency domain encoder, the wavelet transform layer decomposes the preprocessed sample image into four sub-bands: low-frequency smoothing, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency. The edge extraction layer applies an improved Sobel operator to the high-frequency sub-bands to calculate gradients in four directions: 0°, 45°, 90°, and 135°, and fuses them to obtain accurate blood vessel edge features.

[0015] In some embodiments, the feature fusion module includes five feature fusion jump connections. These five feature fusion jump connections use a horizontal-vertical cross-attention mechanism to weightedly fuse corresponding spatial features and frequency domain features, respectively. Specifically, the first spatial feature and the convolutional feature are weightedly fused to obtain a first intermediate feature; the second spatial feature and the first frequency domain feature are weightedly fused to obtain a second intermediate feature; the third spatial feature and the second frequency domain feature are weightedly fused to obtain a third intermediate feature; the fourth spatial feature and the third frequency domain feature are weightedly fused to obtain a fourth intermediate feature; and the fifth spatial feature and the fourth frequency domain feature are weightedly fused to obtain a fifth intermediate feature.

[0016] In some embodiments, the decoding module includes four decoders, each of which includes a convolutional layer and an upsampling layer, wherein: the first decoder combines the fifth intermediate feature and the fourth intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a first decoded feature; the second decoder combines the first decoded feature and the third intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a second decoded feature; the third decoder combines the second decoded feature and the second intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a third decoded feature; the fourth decoder combines the third decoded feature and the first intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a fourth decoded feature.

[0017] In some embodiments, the topology enhancement and segmentation module includes a spatial attention branch, a channel attention branch, and a graph convolutional layer. The spatial attention branch compresses the channel dimension of the fourth decoded feature through a 1×1 convolution to generate a spatial weight map. The channel attention branch generates channel weights through global average pooling and a fully connected layer. The graph convolutional layer captures the topological association between blood vessel pixels, the continuity of blood vessels, and the branching structure by constructing an adjacency matrix. Finally, a pixel-level blood vessel segmentation prediction map is output through a 1×1 convolutional layer and a sigmoid function.

[0018] An embodiment of the second aspect of this application provides a method for efficient segmentation of X-ray blood vessel images, the method comprising the following steps:

[0019] Combined data augmentation preprocessing is performed on the X-ray blood vessel image to be segmented to obtain a preprocessed image;

[0020] The preprocessed image is input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topological relationship enhancement, and the image segmentation result is output.

[0021] The efficient segmentation network model for X-ray blood vessel images based on frequency domain information guidance and topological relationship enhancement is the target segmentation model trained by the training method in any of the above embodiments.

[0022] An embodiment of the third aspect of this application provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, causes the processor to implement the efficient X-ray blood vessel image segmentation method described in the above embodiments.

[0023] In this embodiment, a multi-scale feature extraction layer captures detailed information of different sizes and levels in the image, enabling the network to effectively handle the complexity and irregularity of the blood vessel segmentation region in the image. By combining a wavelet transform module and an edge detection module, the model's ability to perceive blood vessel boundaries is enhanced, improving the accuracy and stability of blood vessel image segmentation. Through topology enhancement and segmentation modules, the network can consider the spatial location between pixels and the correlation between feature channels while focusing on the topological characteristics of blood vessels, accurately capturing the continuity and branching structure of blood vessels, and avoiding excessive reliance on image contrast.

[0024] In this embodiment, the synergistic effect of frequency domain information guidance and topological relationship enhancement effectively solves the segmentation problem of low contrast, complex background and high noise in X-ray vascular images, improves the accuracy and robustness of image segmentation, and can be widely used in the postoperative efficacy evaluation of interventional treatment for diseases such as primary aldosteronism, providing a scientific quantitative basis for clinical diagnosis and treatment. Attached Figure Description

[0025] Figure 1 This is a diagram of an X-ray blood vessel image segmentation model based on frequency domain information guidance and topological relationship enhancement in one embodiment of this application.

[0026] Figure 2 This is a structural diagram of the wavelet transform layer and multi-scale feature extraction layer in one embodiment of this application.

[0027] Figure 3 This is a structural diagram of the edge feature extraction module in one embodiment of this application.

[0028] Figure 4 This is a structural diagram of the topology enhancement and segmentation module in one embodiment of this application. Detailed Implementation

[0029] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings. Preferred embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of this application.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0031] When using the terms “including,” “having,” and “comprising” as described herein, another component may be added unless explicitly qualifying terms such as “only,” “consisting of,” etc. are used. Unless otherwise stated, singular terms may include plural forms and should not be construed as having a quantity of one.

[0032] It should also be understood that, in interpreting an element, although not explicitly described, the element is interpreted as including a range of error, which should be within the acceptable deviation range of a particular value as determined by a person skilled in the art. For example, "approximately," "about," or "substantially" can mean within one or more standard deviations, without limitation herein.

[0033] Traditionally, the assessment of SAAE efficacy relies on multiple clinical indicators (blood biochemistry results). Although numerous preliminary studies have demonstrated the effectiveness and safety of this treatment, a higher level of evidence-based support is still needed before its widespread clinical application. Postoperative imaging changes are a key factor in assessing treatment success and are crucial for judging overall efficacy. Physicians typically rely on experience to observe changes in arterial vascular imaging before and after surgery to judge treatment effectiveness. While this subjective assessment method has some reference value, it also has limitations and uncertainties, making it difficult to provide consistent and accurate evaluation results. Therefore, to improve the accuracy and reliability of the assessment, image analysis combined with deep learning algorithms can better quantify and compare changes before and after surgery, providing a more scientific and reliable basis for efficacy assessment.

[0034] X-ray vessel image segmentation remains extremely challenging due to the influence of various factors, including contrast, complex backgrounds, location and shape, and artifact noise, in pre- and post-operative X-ray adrenal vessel images. Early vessel image segmentation methods primarily relied on optimal thresholding, region growing, filters, and edge detection algorithms. However, these methods demand extremely high image quality and often struggle to achieve accurate segmentation results when processing low-contrast X-ray vessel images with complex backgrounds. Furthermore, these traditional methods often require manual intervention or extensive hyperparameter fine-tuning to achieve satisfactory segmentation results.

[0035] Therefore, this invention provides a model training method, an efficient segmentation method for X-ray vascular images, and a medium to solve the segmentation difficulties caused by low contrast and complex backgrounds in X-ray vascular images. The model training method emphasizes the unique topological characteristics of blood vessels to achieve accurate segmentation; it guides the learning process of vascular features through frequency domain information, improving the accuracy and robustness of X-ray vascular image segmentation and providing a scientific quantitative basis for clinical diagnosis and treatment. It is applicable to the postoperative efficacy evaluation of interventional treatments such as transcatheter ablation of the adrenal artery (SAAE) for primary aldosteronism (PA), and can also be extended to image analysis scenarios for various vascular-related diseases such as coronary heart disease and cerebrovascular diseases.

[0036] Example 1: Model Training

[0037] This embodiment provides a model training method, such as Figures 1 to 4 As shown, Figure 1 This is a diagram of the overall architecture of the segmentation model, used to show the connection relationship and signal flow of the spatial encoder, frequency encoder, feature fusion module, decoding module, topology enhancement and segmentation module, and to demonstrate the structure of the overall segmentation model and the relationships between modules. Figure 2 The structure diagrams of the wavelet transform layer and the multi-scale feature extraction layer are provided, and the parallel structure of multi-scale convolution (conventional convolution and dilated convolution) and the wavelet transform sub-band decomposition process are annotated to clearly present the wavelet transform sub-band decomposition process and the parallel structure of multi-scale convolution. Figure 3 The diagram shows the structure of the edge feature extraction module, illustrating the four-directional gradient calculation and feature weighting fusion process of the improved Sobel operator (edge ​​extraction process and feature processing process). Figure 4 The topology enhancement and segmentation module structure diagram is presented, revealing the synergistic mechanism of spatial attention, channel attention, and graph convolution.

[0038] In this embodiment of the application, the model training method includes the following steps:

[0039] X-ray blood vessel sample images were acquired and combined data augmentation preprocessing was performed to obtain preprocessed sample images;

[0040] The preprocessed sample images are input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology enhancement, and the blood vessel segmentation prediction results are output.

[0041] The standard binary cross-entropy loss function is used as the objective optimization function of the segmentation prediction network to calculate the loss value between the blood vessel segmentation prediction result and the preprocessed sample image label.

[0042] Based on the loss value, the parameters of the efficient X-ray blood vessel image segmentation network model guided by frequency domain information and enhanced by topological relationship are optimized and trained to obtain the target segmentation model.

[0043] In this embodiment, the synergistic effect of frequency domain information guidance and topological relationship enhancement effectively solves the segmentation problem of low contrast, complex background and high noise in X-ray vascular images, improves the accuracy and robustness of segmentation, and can be widely used in the postoperative efficacy evaluation of interventional treatment for diseases such as primary aldosteronism, providing a scientific quantitative basis for clinical diagnosis and treatment.

[0044] In this embodiment, the standard binary cross-entropy loss function is used as the objective optimization function of the segmentation prediction network. The cross-entropy loss function is defined as follows:

[0045]

[0046] in, This represents the index of all pixels in the probability distribution map. It is the first The probability that a pixel belongs to the segmented region It is the first The real label is a pixel.

[0047] In this embodiment, to optimize the model parameters to their optimal state, a stochastic gradient descent (SGD) optimizer and backpropagation algorithm are used to iteratively update the network parameters. The batch size (number of samples used in each iteration) is set to 2, and the initial learning rate is set to 1e-4 to ensure rapid convergence in the early stages of training. To further improve training performance, a LambdaLR learning rate decay technique is employed, dynamically adjusting the learning rate through a custom learning rate scheduling function. This scheduling function can flexibly change the learning rate based on specific metrics or predetermined rules during the training process, thereby providing the most suitable learning rate at different training stages. Furthermore, a momentum term (momentum coefficient set to 0.9) is introduced to accelerate SGD convergence in flat regions and reduce oscillations, ensuring the model can stably reach the optimal solution. With these configurations, the model exhibits good generalization ability and fast convergence speed on large-scale datasets, while reducing the risk of overfitting.

[0048] In some embodiments, the combined data augmentation preprocessing includes: geometric transformation, CLAHE contrast enhancement, Frangi filtering enhancement, and color perturbation. Geometric transformation includes random rotation within a range of -15° to 15°, ±10% pixel translation, and random horizontal / vertical flipping to simulate different imaging angles and body positions. CLAHE contrast enhancement includes using CLAHE (Constrained Contrast Adaptive Histogram Equalization) with a clipping limit of 2.0 and a grid size of 8×8 to improve local contrast. Frangi filtering enhancement includes applying Frangi filtering (σ range 1-3) to highlight vascular tubular structures and suppress background noise. Color perturbation includes randomly adjusting brightness (±15%) and contrast (±10%) to reduce the model's dependence on a fixed color distribution.

[0049] In this embodiment, data augmentation strategies such as random image rotation, random changes to image brightness and contrast, and random changes to image hue and saturation are used to increase sample diversity and ensure the model's generalization ability. Multi-dimensional data augmentation of X-ray vascular images is performed through a combined data augmentation preprocessing method, enriching sample diversity and improving the model's generalization ability.

[0050] In some embodiments, the image resolution is uniformly adjusted to 512×512 pixels after preprocessing, and Z-Score normalization is performed to ensure input consistency.

[0051] In some embodiments, the efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology enhancement includes a feature extraction module, a feature fusion module, a decoding module, and a topology enhancement and segmentation module. The feature extraction module extracts spatial features and frequency domain edge features of the preprocessed sample image in parallel. The feature fusion module performs adaptive weighted fusion of the extracted spatial features and frequency domain edge features. The decoding module combines upsampling and skip connections to restore the resolution of the X-ray blood vessel image. The topology enhancement and segmentation module enhances the topological connectivity of blood vessels through spatial-channel graph convolution and outputs the blood vessel segmentation prediction result.

[0052] In this embodiment, a multi-scale feature extraction layer captures detailed information of different sizes and levels in the image, enabling the network to effectively handle the complexity and irregularity of the blood vessel segmentation region in the image. Combining wavelet transform and edge detection modules enhances the model's ability to perceive blood vessel boundaries, improving the accuracy and stability of blood vessel image segmentation. Through topology enhancement and segmentation modules, the network can consider the spatial location between pixels and the correlation between feature channels while focusing on the topological characteristics of blood vessels, accurately capturing the continuity and branching structure of blood vessels, and avoiding excessive reliance on image contrast.

[0053] In some embodiments, the feature extraction module includes five spatial encoders, one convolutional encoder, and four frequency domain encoders. The five spatial encoders are connected in sequence. The first spatial encoder obtains a first spatial feature with 64 output channels through a convolutional layer and a multi-scale feature extraction layer. The second to fifth spatial encoders each obtain second, third, fourth, and fifth spatial features with 128, 256, 512, and 512 output channels, respectively, through a convolutional layer, a downsampling layer, and a multi-scale feature extraction layer. The convolutional encoder and four frequency domain encoders are connected in sequence. The convolutional encoder consists of two 3×3 convolutions and one 1×1 convolution stacked together, outputting convolutional features. The four frequency domain encoders each include a wavelet transform layer and an edge extraction layer, outputting first, second, third, and fourth frequency domain features, respectively. The convolutional encoder is used to extract preliminary features.

[0054] In some embodiments, the spatial encoder employs a 3×3 convolution and max pooling operation structure in the downsampling layer to reduce the feature map dimension and expand the receptive field. The multi-scale feature extraction layer employs a parallel structure of 3×3 regular convolution, 5×5 convolution, and dilated convolution with dilation coefficients of 2 and 4 to capture local details and global spatial information of blood vessels of different sizes.

[0055] In this embodiment, the spatial encoder utilizes multi-scale convolution operations to capture rich spatial features. In the downsampling layer, convolution and max pooling operations are used to initially extract features from X-ray blood vessel images. This not only helps reduce the spatial dimension of the feature map and improve computational efficiency, but also expands the receptive field and preserves important features, enhancing the model's understanding of image structure and semantic information.

[0056] In this embodiment, the multi-scale feature extraction layer is used to extract features at different scales, enabling the model to adapt to blood vessels of different sizes and improve translation invariance and resistance to deformation. This not only better captures the hierarchical structure and details of X-ray blood vessel images and avoids information loss at a single scale, but also significantly improves the robustness of the model, making it more suitable for processing X-ray blood vessel image data with variable blood vessel regions and irregular shapes.

[0057] In this embodiment, the preprocessed image is used as the input to the network, and a feature vector is generated through the network's forward computation. The network's forward computation includes:

[0058] For convolution operations, the output feature map corresponding to any convolution kernel is calculated using the following formula:

[0059]

[0060] in, Represents a nonlinear excitation function. Indicates the first The bias terms corresponding to each convolution kernel Indicates the input channel index number. Indicates the number of input channels. Indicates the first The first convolutional kernel Channel weight matrix, It's a convolution operation. Indicates the first One input channel image;

[0061] In some embodiments, the frequency domain encoder decomposes the preprocessed sample image into four sub-bands—low-frequency smooth (LL), horizontal high-frequency (LH), vertical high-frequency (HL), and diagonal high-frequency (HH)—through a wavelet transform layer to achieve separation of high and low frequency information. The edge extraction layer applies an improved Sobel operator to the high-frequency sub-bands (LH, HL, HH) to calculate gradients in four directions: 0°, 45°, 90°, and 135°, and then fuses them to obtain accurate blood vessel edge features.

[0062] In this embodiment, the frequency domain encoder extracts vascular boundary features through wavelet transform, convolution operation, and the Sobel edge detection operator, enhancing the extraction of detailed information from vascular images. In the wavelet transform part, the input features are decomposed into four sub-bands, representing high and low frequency information in different directions. Each sub-band has half the width and height of its original size, equivalent to a downsampling operation. This process effectively separates high-frequency (edges and textures) and low-frequency (smooth regions) information in X-ray vascular images while preserving key details. For the edge feature extraction part, based on the wavelet transform, the Sobel edge detection operator is applied to the high-frequency sub-band information to accurately capture the edge features of the blood vessels.

[0063] In wavelet transform operations, the Discrete Wavelet Transform (DWT) is a mathematical tool for decomposing a signal into different frequency subbands. In two-dimensional image processing, DWT typically decomposes the input image into four subbands, each representing a different frequency component. For two-dimensional discrete wavelet transform, the input image f(x,y) is convolved by a low-pass filter h(n) and a high-pass filter g(n), and then downsampled to obtain four subbands. This process can be represented by the following unified formula:

[0064]

[0065] in, d1,d2 represents the subband coefficients after the (j+1)th level decomposition. (d1,d2) can take the values ​​(L,L), (L,H), (H,L), (H,H), which correspond to the LL, LH, HL and HH subbands, respectively. and It is a filter that varies depending on the choice of d1 and d2. When d1 or d2 is L, the corresponding filter is a low-pass filter h(n); when d1 or d2 is H, the corresponding filter is a high-pass filter g(n). m and n are the indices of the output subband, and k and l traverse all pixel positions of the input image f(x,y). k and l are the two-dimensional index variables of the image pixel traversal.

[0066] In edge detection, the Sobel operator is a classic edge detection operator used to extract edge information in image processing. It identifies regions with significant gray-level changes (i.e., edges) by calculating the gray-level gradient of the image. An improved Sobel operator is used to extract gradients in the 0°, 45°, 90°, and 135° directions of the image, expressed by the following formula:

[0067]

[0068]

[0069] Where P represents the image, * is the convolution operator, and G is the output of the edge detection;

[0070] In this embodiment, dedicated convolution templates for two diagonal directions, 45° and 135°, are added, and operator weights are redistributed to achieve synchronous gradient calculation in four directions, covering horizontal, vertical, and two diagonal edge information. This captures more diagonal image edges, avoids missing diagonal edges, and makes the final extracted edge details richer and more continuous.

[0071] In some embodiments, the feature fusion module includes five feature fusion jump connections. These five feature fusion jump connections use a horizontal-vertical cross-attention mechanism to weight and fuse corresponding spatial features and frequency domain features respectively. The first spatial feature and the convolutional feature are weighted and fused to obtain a first intermediate feature; the second spatial feature and the first frequency domain feature are weighted and fused to obtain a second intermediate feature; the third spatial feature and the second frequency domain feature are weighted and fused to obtain a third intermediate feature; the fourth spatial feature and the third frequency domain feature are weighted and fused to obtain a fourth intermediate feature; and the fifth spatial feature and the fourth frequency domain feature are weighted and fused to obtain a fifth intermediate feature.

[0072] In this embodiment, the feature fusion module generates query and key vectors in the horizontal and vertical directions through a horizontal-vertical cross-attention mechanism. It captures the correlation between spatial features and frequency domain features through attention calculation, thereby achieving complementarity between the global semantic information of spatial features and the edge detail information of frequency domain features, and generating a weighted fusion result of spatial features and frequency domain edge features.

[0073] In this embodiment, a feature fusion module effectively fuses multi-scale features of extracted local and global information, enhancing the model's ability to perceive details of vascular structures. This fusion not only preserves the spatial information of the original image but also further improves the accuracy and robustness of segmentation by introducing frequency domain features and edge information.

[0074] In this embodiment, by using skip connections to fuse feature maps of corresponding layers of the spatial encoder, the features of low-level and high-level layers are combined, which can achieve key functions such as information transmission, detail preservation, anti-gradient vanishing, accelerated convergence, and reduced parameter transmission distance. This can enhance the preservation of detailed information, ensure that key features such as small branches of blood vessels are not missed, and significantly improve network performance and generalization ability.

[0075] In some embodiments, the decoding module includes four decoders, each of which includes a convolutional layer and an upsampling layer; the first decoder combines the fifth intermediate feature and the fourth intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a first decoded feature; the second decoder combines the first decoded feature and the third intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a second decoded feature; the third decoder combines the second decoded feature and the second intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a third decoded feature; the fourth decoder combines the third decoded feature and the first intermediate feature, and then performs feature mapping through the convolutional layer and the upsampling layer to obtain a fourth decoded feature.

[0076] In this embodiment, the main task of the decoder is to convert the high-level semantic feature maps extracted by the encoder into an output with the same resolution as the input image, and to gradually restore the detailed information. The decoder adopts a structure of linear interpolation upsampling, convolutional layers, BatchNorm (Batch Normalization) combined with ReLU (Rectified Linear Unit) to gradually restore the fused feature map to the resolution of the input image.

[0077] In this embodiment, before each upsampling stage, the features of the current layer of the decoder are added to the feature map of the corresponding layer of the encoder to form an enhanced feature map. This fusion step ensures the full combination of high-frequency details and semantic information, and gradually converts the added feature map into a high-resolution feature map.

[0078] In this embodiment, batch normalization (BN) is a method to make neural network training faster and more stable. It calculates the mean and variance of each mini-batch and pulls them back to a standard normal distribution with a mean of 0 and a variance of 1. Layer normalization (LN) is a neural network regularization method, similar to batch normalization. It adjusts the mean and variance of all features of each sample in the same layer to a standard normal distribution to improve training speed and stability.

[0079] In this embodiment, the decoder further includes nonlinear excitation and probability value transformation. The nonlinear excitation uses a rectified linear unit (ReLU) as the nonlinear excitation function, performing a nonlinear transformation on each normalized value. The rectified linear unit (ReLU) is defined as follows:

[0080]

[0081] Here, max represents finding the maximum value, and x is an input value.

[0082] The probability value transformation uses the Sigmoid function to convert the predicted score into a probability distribution. The Sigmoid function is defined as follows:

[0083]

[0084] in, It is the probability that the j-th pixel belongs to the segmented region. It is the prediction score at the j-th pixel, which is the final output of the label predictor.

[0085] In this embodiment, during the decoder stage, transposed convolution is used to restore the image size, while skip connections compensate for any potentially lost information, helping to recover key features of the vascular structure. Subsequently, spatial features and frequency domain edge features are fused through operations such as convolution, enhancing the model's perception of detailed information and improving segmentation accuracy. Through this process, the decoder not only restores the image resolution but also effectively preserves low-level spatial details and high-level semantic information, enabling the network to generate fine and accurate segmentation results, making it particularly suitable for processing complex X-ray vascular images.

[0086] In some embodiments, the topology enhancement and segmentation module includes a spatial attention branch, a channel attention branch, and a graph convolutional layer. The spatial attention branch compresses the channel dimension of the fourth decoded feature through a 1×1 convolution to generate a spatial weight map that highlights the vascular region. The channel attention branch generates channel weights through global average pooling and a fully connected layer to enhance effective feature channels. The graph convolutional layer captures the topological associations between vascular pixels, the continuity of vascular vessels, and the branching structure by constructing an adjacency matrix. Finally, it outputs a pixel-level vascular segmentation prediction map through a 1×1 convolutional layer and a sigmoid function.

[0087] In this embodiment, before outputting the segmentation prediction map, a topology enhancement and segmentation module is introduced to strengthen the topological connectivity of blood vessels, so as to ensure the accuracy and connectivity of blood vessel structures and achieve efficient segmentation of X-ray blood vessel images.

[0088] An embodiment of the second aspect of this application provides a method for efficient segmentation of X-ray blood vessel images, the method comprising the following steps:

[0089] Combined data augmentation preprocessing is performed on the X-ray blood vessel image to be segmented to obtain a preprocessed image;

[0090] The preprocessed image is input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topological relationship enhancement, and the image segmentation result is output.

[0091] The efficient segmentation network model for X-ray blood vessel images based on frequency domain information guidance and topological relationship enhancement is the target segmentation model trained by the training method in any of the above embodiments.

[0092] An embodiment of the third aspect of this application provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, causes the processor to implement the efficient X-ray blood vessel image segmentation method described in the above embodiments.

[0093] In this embodiment, a multi-scale feature extraction layer captures detailed information of different sizes and levels in the image, enabling the network to effectively handle the complexity and irregularity of the blood vessel segmentation region in the image. By combining a wavelet transform module and an edge detection module, the model's ability to perceive blood vessel boundaries is enhanced, improving the accuracy and stability of blood vessel image segmentation. Through topology enhancement and segmentation modules, the network can consider the spatial location between pixels and the correlation between feature channels while focusing on the topological characteristics of blood vessels, accurately capturing the continuity and branching structure of blood vessels, and avoiding excessive reliance on image contrast.

[0094] In one example, the Kaiming initialization method is used to initialize the network parameters, with the binary cross-entropy loss function as the optimization objective. The parameters are iteratively updated using a stochastic gradient descent (SGD) optimizer combined with a LambdaLR learning rate decay strategy. Specifically, this includes:

[0095] (1) Parameter initialization: The Kaiming initialization method is used to initialize the parameters of all convolutional and fully connected layers of the network, and to initialize the parameters of the X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology relationship enhancement.

[0096] (2) Dataset construction: The dataset is from Daping Hospital. A total of 216 X-ray angiography images of 108 patients with primary aldosteronism before and after ablation surgery were used. The images were stratified at a ratio of 3:7 to form a training set (60 images) and a test set (156 images). All images were pixel-level annotated by senior physicians using the Labelme tool to ensure the accuracy and professionalism of the data.

[0097] In the above operation, in order to evaluate the model performance, a stratified sampling method was used. 60 X-ray angiography images of 30 patients were randomly selected as the training set, while 156 images of the remaining 78 patients were assigned to the test set to comprehensively test the model's generalization ability.

[0098] The regions of interest in the dataset were precisely divided, and the resolution was then uniformly adjusted to 512×512 pixels to ensure input consistency. To enrich the diversity of samples and improve the robustness of the model, various data augmentation strategies were employed, including combined data augmentation for all images: random rotation from -15° to 15°, ±10% pixel translation, random flipping, CLAHE filtering (clipping limit 2.0, 8×8 grid), Frangi filtering (σ=2), and brightness ±15% and contrast ±10% adjustments. CLAHE (Contrast-Limited Adaptive Histogram Equalization) and Frangi filtering were used to enhance the original images, improving image contrast and highlighting key details. Various geometric transformations were applied, such as random rotation, translation, and random horizontal and vertical flipping within the range of -15 to 15 degrees, to simulate different imaging angles and positional changes. This approach effectively prevented overfitting while expanding the data volume.

[0099] (3) Loss function and optimizer: The binary cross-entropy loss function is adopted, combined with the stochastic gradient descent (SGD) optimizer, the batch size is set to 2, the initial learning rate is 1e-4, the momentum coefficient of the SGD optimizer is 0.9, and the weight decay coefficient is 1e-5.

[0100] Learning rate scheduling: The LambdaLR strategy is adopted. The training takes 100 rounds. The learning rate is kept constant for the first 50 rounds, and then decays by 0.9 every 10 rounds thereafter.

[0101] An early stopping strategy is adopted, using the test set dice coefficient Dice as the monitoring metric. Training is stopped when there is no improvement for 15 consecutive rounds, and the optimal model is saved.

[0102] The network input consists of data-enhanced X-ray images of blood vessels, measuring 512×512 pixels with 3 channels. Two images are input at a time, with dimensions [2,3,512,512]. The network output consists of 1 channel, with the same size as the input images and dimensions [2,1,512,512].

[0103] The original input is mapped to a high dimension before an encoder is used to extract feature information. The number of channels in the output feature of each encoder is 64, 128, 256, 512, and 512, respectively. The decoder is constructed from four simple convolutional layers, followed by a spatial-channel attention module to enhance the vascular topology features. Then, convolutional layers are used to restore the original image size and predict the output segmentation prediction map. Detailed model parameters are shown in the table below.

[0104] Table 1. Parameters of X-ray blood vessel image segmentation network based on frequency domain information guidance and topology enhancement.

[0105]

[0106] Table 2. Parameter table of multi-scale feature extraction layer (taking feature extraction layer 1 as an example)

[0107]

[0108] Table 3 Wavelet transform layer parameters (taking wavelet transform layer 1 as an example)

[0109]

[0110] Table 4. Parameters of the edge extraction layer (taking edge extraction layer 1 as an example)

[0111]

[0112] Table 5. Parameters of the feature fusion layer (taking feature extraction layer 1 as an example)

[0113]

[0114] Table 6. Convolutional layer parameters

[0115]

[0116] In this embodiment, model testing and performance verification include: inputting test set images into the trained model, and evaluating performance using sensitivity (SE), precision (Precision), accuracy (ACC), intersection-over-union (IoU), and dice coefficient (Dice). The results are shown in the table below:

[0117] Table 7 Performance Evaluation Table

[0118]

[0119] As shown in Table 7, our method is significantly better than the comparison algorithm in all indicators, with the Dice coefficient improved by 6.37-10.66 percentage points, and it can accurately segment vascular structures in low-contrast and complex backgrounds.

[0120] The method described in this application was applied to evaluate the efficacy of SAAE in 15 PA patients to verify its clinical application. The change in vascular cross-sectional area was calculated through the segmentation results. The change in vascular cross-sectional area was found to be positively correlated with clinical blood biochemistry test (aldosterone level) (correlation coefficient 0.93). The intragroup correlation coefficient (ICC) between the segmentation results and the physician's manually labeled results reached 0.93, which meets the clinical consistency requirements and can provide an objective quantitative basis for postoperative efficacy evaluation.

[0121] In some embodiments, five widely used metrics are employed to evaluate the accuracy performance of efficient X-ray vessel image segmentation methods, including sensitivity (SE), precision (Precision), accuracy (ACC), intersection-to-union ratio (IOU), and dice coefficient (Dice), defined as follows:

[0122]

[0123]

[0124]

[0125]

[0126]

[0127] Where TP (True Positive) and TN (True Negative) represent the number of correctly segmented blood vessel pixels and background pixels, respectively; FP (False Positive) represents background pixels that are incorrectly labeled as blood vessel pixels; and FN (False Negative) represents blood vessel pixels that are incorrectly predicted as background pixels.

[0128] 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.

[0129] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A model training method, characterized in that, Includes the following steps: X-ray blood vessel sample images were acquired and combined data augmentation preprocessing was performed to obtain preprocessed sample images; The preprocessed sample images are input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology enhancement, and the blood vessel segmentation prediction results are output. The standard binary cross-entropy loss function is used as the objective optimization function of the segmentation prediction network to calculate the loss value between the blood vessel segmentation prediction result and the preprocessed sample image label. Based on the loss value, the parameters of the efficient X-ray blood vessel image segmentation network model guided by frequency domain information and enhanced by topological relationship are optimized and trained to obtain the target segmentation model.

2. The model training method according to claim 1, characterized in that, Efficient X-ray blood vessel image segmentation network models based on frequency domain information guidance and topology enhancement include: The feature extraction module extracts spatial features and frequency domain edge features of the preprocessed sample images in parallel. The feature fusion module adaptively weights and fuses the extracted spatial features with frequency domain edge features; The decoding module combines upsampling and skip connections to restore the resolution of X-ray vascular images; The topology enhancement and segmentation module enhances the topological connectivity of blood vessels through spatial-channel graph convolution and outputs the blood vessel segmentation prediction results.

3. The model training method according to claim 2, characterized in that, The feature extraction module includes five spatial encoders, one convolutional encoder, and four frequency domain encoders; Five spatial encoders are connected in sequence. The first spatial encoder obtains the first spatial feature with 64 output channels through a convolutional layer and a multi-scale feature extraction layer. The second to fifth spatial encoders obtain the second, third, fourth and fifth spatial features with 128, 256, 512 and 512 output channels respectively through a convolutional layer, a downsampling layer and a multi-scale feature extraction layer. A convolutional encoder and four frequency domain encoders are connected in sequence. The convolutional encoder consists of two stacked 3×3 convolutions and one 1×1 convolution, and outputs convolutional features. The four frequency domain encoders each include a wavelet transform layer and an edge extraction layer, and output the first frequency domain feature, the second frequency domain feature, the third frequency domain feature and the fourth frequency domain feature in sequence.

4. The model training method according to claim 3, characterized in that, In the spatial encoder, the downsampling layer adopts a 3×3 convolution and max pooling operation structure, and the multi-scale feature extraction layer adopts a parallel structure of 3×3 regular convolution, 5×5 convolution and dilated convolution with dilation coefficients of 2 and 4.

5. The model training method according to claim 3, characterized in that, In the frequency domain encoder, the wavelet transform layer decomposes the preprocessed sample image into four sub-bands: low-frequency smoothing, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency. The edge extraction layer applies an improved Sobel operator to the high-frequency sub-bands to calculate gradients in four directions: 0°, 45°, 90°, and 135°, and then fuses them to obtain accurate blood vessel edge features.

6. The model training method according to claim 2, characterized in that, The feature fusion module includes five feature fusion hop connections. These five connections use a horizontal-vertical cross-attention mechanism to weight and fuse the corresponding spatial and frequency domain features, respectively. The first intermediate feature is obtained by weighted fusion of the first spatial feature and the convolutional feature; The second intermediate feature is obtained by weighted fusion of the second spatial feature and the first frequency domain feature; The third intermediate feature is obtained by weighted fusion of the third spatial feature and the second frequency domain feature; The fourth intermediate feature is obtained by weighted fusion of the fourth spatial feature and the third frequency domain feature; The fifth spatial feature and the fourth frequency domain feature are weighted and fused to obtain the fifth intermediate feature.

7. The model training method according to claim 2, characterized in that, The decoding module includes four decoders, each of which includes a convolutional layer and an upsampling layer, wherein: The first decoder combines the fifth and fourth intermediate features, and then performs feature mapping through convolutional and upsampling layers to obtain the first decoded feature; The second decoder combines the first decoded feature and the third intermediate feature, and then performs feature mapping through convolutional layers and upsampling layers to obtain the second decoded feature; The third decoder combines the second decoded feature and the second intermediate feature, and then performs feature mapping through convolutional layers and upsampling layers to obtain the third decoded feature; The fourth decoder combines the third decoded feature and the first intermediate feature, and then performs feature mapping through convolutional layers and upsampling layers to obtain the fourth decoded feature.

8. The model training method according to claim 2, characterized in that, The topology enhancement and segmentation module includes a spatial attention branch, a channel attention branch, and a graph convolutional layer. The spatial attention branch compresses the channel dimension of the fourth decoded feature through a 1×1 convolution to generate a spatial weight map. The channel attention branch generates channel weights through global average pooling and a fully connected layer. The graph convolutional layer captures the topological relationships between blood vessel pixels, the continuity of blood vessels, and the branching structure by constructing an adjacency matrix. Finally, a pixel-level blood vessel segmentation prediction map is output through a 1×1 convolutional layer and a sigmoid function.

9. A method for efficient segmentation of X-ray blood vessel images, characterized in that, Includes the following steps: Combined data augmentation preprocessing is performed on the X-ray blood vessel image to be segmented to obtain a preprocessed image; The preprocessed image is input into an efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topological relationship enhancement, and the image segmentation result is output. The efficient X-ray blood vessel image segmentation network model based on frequency domain information guidance and topology relationship enhancement is the target segmentation model trained by the training method described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, the processor enables the processor to implement the efficient segmentation method for X-ray blood vessel images as described in claim 9.

Images