A fourier frequency domain enhanced blood vessel segmentation method based on spatial topology perception

Abstract

The application discloses a Fourier frequency domain enhanced blood vessel segmentation method based on space topology perception, which comprises the following steps: collecting detailed CT image data and performing multidimensional pretreatment, designing an encoder and a decoder, constructing a frequency domain enhanced network architecture, establishing a blood vessel dissection dependent model, learning blood vessel levels from multiple angles in topology, performing attention interaction from bottom to top, and performing multi-stage training and verifying experimental results; through a Fourier coding module and a Fourier fusion decoding module, the image is converted from a spatial domain to a frequency domain, the characteristics of blood vessel structure, low frequency and noise high frequency are utilized to fuse the high-frequency edge features of the encoder with the low-frequency semantic features of the decoder, the influence of noise is reduced, the segmentation precision of small blood vessels and low-contrast regions is improved, the problems of false segmentation and missed segmentation caused by image noise are solved, and compared with a traditional convolutional network in a spatial domain, the robustness to noise and segmentation accuracy are significantly improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of deep learning and medical image segmentation, specifically a Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness. Background Technology

[0002] Medical imaging technology is crucial for preoperative surgical planning. Among them, magnetic resonance imaging (MRI) and computed tomography (CT) are widely used. Although MRI is favored due to its lack of radiation, it has limitations such as time-consuming examination, high cost, and high requirements for patients. CT, on the other hand, dominates in routine clinical examinations due to its advantages of low cost, fast scanning, and high spatial resolution. The anatomical information of organs and blood vessels it provides lays the foundation for segmentation tasks. However, current vascular segmentation relies heavily on the doctor's experience, which is not only inefficient but also easily affected by subjective factors. Furthermore, the complex structure of blood vessels and their low contrast with surrounding tissues lead to frequent missegmentation and omissions. Therefore, the development of automated vascular segmentation technology has great clinical value and scientific research significance. In the field of medical image segmentation, commonly used methods are divided into two categories: traditional image processing and deep learning. Traditional methods, such as thresholding and edge detection, are based on image features such as brightness and color. They have the advantages of being simple, fast, and interpretable, but their accuracy and robustness are poor when processing complex images. For example, the Vesseness filter improved by Cui et al., although optimized for coronary artery segmentation, still has limitations. In recent years, deep learning has driven innovation in vascular segmentation technology. The multi-stage 2D convolutional neural network algorithm proposed by Xu et al. improves the accuracy of pulmonary vascular segmentation through step-by-step processing. Yan et al.'s LVSNet uses innovative modules to achieve layered segmentation of liver vessels. Huang et al. solved the problem of imbalanced liver vascular data based on 3DU-Net. Tong et al. improved the network by reducing false positive areas in hepatic vein segmentation. Although these advances have significantly improved the segmentation effect, they still have many shortcomings: Significant image noise interference: During CT scanning, factors such as limitations in the hardware performance of imaging equipment, physiological movements such as patient breathing or body movement, and improper scanning parameter settings can all introduce noise into vascular images. These noises are similar to the structural features of blood vessels, making it difficult for segmentation algorithms to effectively distinguish between the two. Noise interference is particularly prominent in small blood vessels or areas with low contrast, which can easily lead to missegmentation or missed segmentation of blood vessels. Severe image contrast deficiency: Blood vessels and surrounding tissues exhibit low contrast in CT images, especially in small blood vessels and their branching areas, making boundary identification extremely difficult. On the one hand, the density difference between the blood flowing inside the blood vessel and the surrounding tissue is small, directly resulting in low image contrast. On the other hand, as the diameter of the blood vessel decreases, the blood flow velocity increases, further weakening the contrast with the surrounding tissue. This low contrast problem is particularly evident in the bifurcation and intersection areas of blood vessels. The complex interaction between the vascular structure and tissue further exacerbates the difficulty of blood vessel segmentation. Vascular size varies greatly: Blood vessels vary significantly in size and shape between different individuals and in different parts of the same organism. For example, in cerebral angiography, the diameter of small arteries and microvessels is only a few micrometers, while the diameter of the main cerebral arteries can reach several millimeters. Such a huge size range places stringent requirements on segmentation algorithms. The algorithm must be able to adapt to the characteristics of blood vessels at different scales at the same time in order to accurately identify and segment blood vessel structures of various sizes. The vascular structure is intricate and complex: the vascular network is composed of a large number of branches, intersections and small vessels, with extremely complex morphology and distribution. The vascular branching is rich in layers, subdividing from the main vessels to the capillaries. The branches vary in length and thickness, and the shape and angle of the vascular intersections are diverse. This complex vascular topology greatly increases the technical difficulty of vascular segmentation. Summary of the Invention

[0003] The purpose of this invention is to provide a Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness, in order to solve the problems mentioned in the background art, such as significant image noise interference, severe image contrast deficiency, huge differences in blood vessel size, and complex blood vessel structure.

[0004] Therefore, the present invention provides a Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness, comprising the following steps: S1. Acquire detailed CT image data and perform multi-dimensional preprocessing; S2. Design the encoder and decoder, and build a frequency domain enhancement network architecture; S3. Establish a vascular anatomy-dependent model; S4, Multi-angle topological learning of blood vessel layers; S5. Bottom-up attention interaction; S6. Conduct multi-stage training and verify the experimental results.

[0005] Preferably, in step S1, the specific steps for acquiring detailed CT image data and performing multi-dimensional preprocessing are as follows: collect abdominal and lung CT images, have clinical experts annotate vascular masks using Mimics to cover different types of blood vessels in the hepatic artery and vein and pulmonary artery, then crop abnormal intensity values ​​of CT images, normalize the intensity using Z-score, and then enhance composite data, including using random Gaussian noise, random contrast enhancement, random mirror flip, random horizontal flip, and random rotation.

[0006] Preferably, in step S2, the specific steps for constructing the frequency domain enhancement network architecture and designing the encoder are as follows: (1) First, input a 3D image, extract initial features through large kernel convolution, and use these features as input to the encoder. Then, use DFT and IDFT to transform the features between the spatial and frequency domains: where Used to fill the frequency representation, in order to prevent the wrapping effect in subsequent convolution operations, the frequency domain representation is padded in advance. The zero-filling sequence can provide enough "space" for the linear convolution of the sequence to prevent the tail of a sequence from wrapping and interfering with the beginning of the convolution. (2) express Image spatial representation of layers After normalization by the LN layer, the signal is then subjected to Discrete Fourier Transform (DFT). DFT can decompose complex signals into single frequency components with different amplitudes, thereby enabling operations such as image filtering. (3), formula In the middle, frequency is represented With learnable weights Element-wise multiplication is performed between them, similar to convolution in the spatial domain. The result of the multiplication is cropped to remove the padding region and a frequency domain learnable bias term is added. (4), formula In this context, the inverse discrete Fourier transform is used to represent the frequency. The data is transformed back to the spatial domain, and after being normalized through an LN layer, it is connected to a multilayer perceptron (MLP) and then added to the spatial data of the previous layer. This forms a residual, i.e., a jump connection. The formula represents the final spatial data. By using pooling operations to downsample the spatial representation, the Fourier transformer coding block effectively integrates traditional convolutional neural network operations with frequency domain processing techniques, which can fully utilize the advantages of both representations and improve network performance.

[0007] Preferably, in step S2, the specific steps for constructing the frequency domain enhancement network architecture and designing the decoder are as follows: the specific formula for the Fourier fusion decoder is... The features derived from the encoder are denoted as E, and the features derived from the decoder are denoted as D, with superscripts... and These represent the layer depths, with the subscripts 'spatial' and 'freq' used to distinguish between spatial and frequency domain features after Fourier transform. Then, a Cropinner is applied to the shallow features of the encoder, and a Cropouter is applied to the deep features of the decoder, fusing low-frequency semantics with high-frequency details. Finally, the spatial domain features are transformed using the Inverse Discrete Fourier Transform (IDFT). The transformation is used as the input to the next layer of the decoder.

[0008] Preferably, in step S3, the specific steps for establishing the vascular anatomy-dependent model are as follows: (1) Use the constructed FDESNet network to perform preliminary reasoning on CT images, generate vascular mask prediction results, locate the approximate area of ​​the blood vessels, and then use a 5×5×5 three-dimensional expansion kernel to expand the labeled real vascular mask to generate DilatedLabel, expand the coverage of the vascular area, and ensure that small branch voxels are included. (2) Based on the expanded DilatedLabel, calculate its minimum outer rectangle in three-dimensional space, determine the spatial range containing all expanded blood vessel voxels, and then use the above rectangle to crop the original CT image, remove irrelevant background areas, reduce noise interference, and finally the cropped image size dynamically adapts to the blood vessel distribution. For example, liver blood vessels can be cropped to 128×128×64 voxels. (3) Extract the blood vessel and organ point set. Extract the three-dimensional coordinates of all blood vessel voxels from the expanded blood vessel mask to form the blood vessel point set. , dimension ,in Point sets are also extracted from adjacent organs of blood vessels. Then, the farthest point sampling method is used to downsample the point cloud on the organ surface, retaining 8000 sample points to reduce the computational complexity of the subsequent calculations. (4) Normalize and crop the distance field map to reduce the interference of outliers. Calculate the Euclidean distance from each point in the blood vessel point set to all points in the organ point set, and take the minimum value to generate the original distance field. Then map the original distance values ​​to the [0,1] interval and set a threshold to crop the distance values ​​that are out of range to the threshold boundary to reduce the interference of outliers. (5) Finally, the standardized distance field map is combined with the cropped CT image and the vascular mask prediction results as anatomical dependent context information and input into the subsequent topology learning module. By using the cropped small-size image and downsampled organ point set, the computational complexity of the distance field is reduced, the training efficiency is improved, and the anatomical dependent information is integrated.

[0009] Preferably, in step S4, the specific steps for multi-angle topological learning of blood vessel layers are as follows: (1) The distance field map generated in the first part and the cropped CT image region are used as input. Then, the CT image is cropped by the outer rectangular box of the expanding blood vessel mask to reduce background noise and accelerate training. FDESNet is used as the backbone network and Fourier Transformer encoder is combined to generate multi-task features. The generated features are divided into three task branches, which are used for key point detection, centerline heat map regression and adjacent connection prediction, respectively. (2) For the blood vessel key point detection branch, a probability map of each point as a key point is generated by FDESNet. The probability map represents the probability value of each point being predicted as a key point. Then, the weighted Hausdorff distance between the predicted point and the ground real point is calculated to calculate their similarity. (3) For the centerline heatmap regression branch, a centerline heatmap unit is introduced, and a local normalized heatmap is generated by combining it with a 3D adaptive Gaussian filter. Defined as: and ,in These are the coordinates of point t on the center line. It means to pass through and the first boundary point on the cross section perpendicular to the centerline coordinates, yes The width of the core is used as the radius of the blood vessel, and finally the shape of the heat map is dynamically adjusted based on the diameter of the blood vessel to ensure the consistency of the skeleton line; (4) For the adjacent connection prediction branch, taking the 26 nearest neighbors in the 3×3×3 cube region as the object, given the blood vessel label And its 26 nearest neighbors , will neighboring Index from 1 to 26, and pair each blood vessel. A binary tag encoded as a 1-channel label, if and If they are all blood vessel points, then they should be... Marked as a connection, corresponding Points in the channel It is marked as 1 if it is not connected, otherwise it is marked as 0, indicating no connection. Then training and prediction are performed. During the training phase, points are generated. and 26 nearest neighbors The prediction vectors of a total of 27 channels are generated, and then a channel probability map with connectivity likelihood is generated. In the prediction stage, the channel probability map is converted into a binary image as the final segmentation result, and pairs of voxels remain consistent with each other in terms of connectivity.

[0010] Preferably, in step S5, the specific steps for bottom-up attention interaction are as follows: (1) The key point detection probability map, centerline heatmap and adjacent connection probability map output by the three branches are spliced ​​along the channel dimension. The generated comprehensive feature map is subjected to global average pooling and full connection operation. The importance weight of each channel is calculated. Then, based on the channel attention weighted features, the attention weight of the spatial dimension is calculated to focus on the key regions of vascular topology. (2) Multiply the channel attention and spatial attention to obtain the final attention weight, weight the spliced ​​comprehensive feature map, and then reduce the dimensionality of the weighted comprehensive feature map through a three-dimensional convolutional layer to generate a feature map adapted to the segmentation task. (3) Generate a segmentation mask and perform post-processing. Input the feature map adapted to the segmentation task into the sigmoid activation function to generate a blood vessel segmentation probability map. Then, use an adaptive threshold to convert the probability map P into a binary mask M. Based on the topological constraints of the adjacent connection prediction branches, perform morphological operations on the binary mask to ensure the continuity of blood vessel connections. (4) Finally, a three-dimensional Fourier transform is performed on the cropped image region to extract frequency domain features. The frequency domain information is embedded into the spatial features through a Fourier transform encoder. The features embedded with frequency domain information are weighted and fused with spatial attention features to enhance the segmentation accuracy of blood vessel details.

[0011] Preferably, in step S6, the specific steps for performing multi-stage training and verifying experimental results are as follows: (1) Perform phased progressive training. Use the expanded blood vessel mask annotation data to train the FDESNet network to initially locate the blood vessel region. At the same time, generate a distance field map to encode the spatial relationship between blood vessels and adjacent organs. Reduce background noise interference by cropping CT images. Then input the distance field map and the cropped image. Simultaneously train the three branches of key point detection, centerline heat map regression, and adjacent connection prediction. Use the Fourier Transformer encoder and attention mechanism to enhance the interaction of different topological features. Finally, connect the anatomical dependency module and the topology learning module, add bottom-up attention interaction, and fine-tune the overall parameters to ensure the continuity of the model from coarse segmentation to fine topology learning. (2) Divide the public datasets such as 3Dircadb into training set, validation set and test set in an 8:1:1 ratio to ensure that each subset contains the vascular morphology of different individuals. Then normalize the intensity of the CT images, crop outliers and scale them to a uniform range to eliminate the influence of equipment differences. Enhance the data by randomly rotating, flipping and scaling to enhance the model’s adaptability to blood vessels at different angles. Add Gaussian noise and adjust the contrast to simulate the noise scene in clinical scanning. Expand the real mask of blood vessels to expand the annotation range to cover small branches. (3) Calculate the overlap rate between the predicted mask and the real label, measure the segmentation accuracy, count the proportion of correctly segmented pixels, evaluate the model's ability to distinguish between the background and blood vessels, calculate the proportion of correctly segmented real blood vessels to reflect the ability to capture small blood vessels, measure the proportion of the intersection of the predicted and real regions to the union, supplement the evaluation dimensions of the Dice index, the benchmark methods are 3DU-Net, DV-Net and nnUNet, the same data preprocessing process and training rounds are used, the same hardware environment and loss function configuration are used uniformly to ensure the credibility of the comparison results; (4) The results show that the method of the present invention has achieved the best segmentation performance, with the highest values ​​of 72.34%, 99.47%, 73.12% and 56.79% in Dice, accuracy ACC, sensitivity Sen and Jaccard index respectively, which confirms the superiority of the Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness.

[0012] The present invention proposes a Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness, the advantages of which are as follows: This Fourier frequency domain enhanced blood vessel segmentation method transforms the image from the spatial domain to the frequency domain through a Fourier Transformer encoding module and a Fourier fusion decoding module. Utilizing the low-frequency characteristics of blood vessel structures and high-frequency noise, a frequency domain filter is designed to suppress high-frequency noise while preserving low-frequency blood vessel information. By fusing high-frequency edge features from the encoder with low-frequency semantic features from the decoder, the method reduces the impact of noise while improving the segmentation accuracy for small blood vessels and low-contrast regions. This solves the problems of missegmentation and missed segmentation caused by image noise. Compared with traditional spatial domain convolutional networks, it significantly improves robustness to noise and segmentation accuracy. This Fourier frequency domain enhanced blood vessel segmentation method utilizes anatomical structure to enhance the recognition of low-contrast blood vessels. It incorporates a blood vessel anatomy-dependent stage during training to generate a distance field map that encodes the distance relationship between blood vessels and surrounding organs, providing accurate information on blood vessel location and spatial dependence. Even in regions with small blood vessel diameters and low contrast, the model can improve segmentation accuracy by leveraging this structural dependence information, reducing missegmentation or missed segmentation caused by low contrast, thus making up for the shortcomings of existing technologies in low-contrast blood vessel recognition. In the vascular topology learning stage of this Fourier frequency domain enhanced vascular segmentation method, a multi-task learning strategy is adopted. Through key point detection, centerline heatmap regression, and adjacent connectivity prediction, the multi-level topological structure of blood vessels is captured. Key point detection identifies bifurcation and terminal points, centerline heatmap regression maintains the vascular skeleton structure and dynamically adapts to changes in blood vessel diameter, and adjacent connectivity prediction describes the vascular context. This approach enables the model to adapt to huge changes in blood vessel size from small branches to main vessels, effectively handling the problems of complex vascular structures, numerous branches, and varied intersection morphologies. Compared with traditional skeleton line extraction methods, it can learn vascular structure information more comprehensively and accurately. The bottom-up attention interaction module of this Fourier frequency domain enhanced blood vessel segmentation method uses a self-attention mechanism to transfer the adjacency connectivity information of dense layers to the centerline and key points of sparse layers. This effectively integrates features from different levels and avoids the problem that existing attention mechanisms cannot effectively integrate cross-level features and lose spatial hierarchical structure. This allows the model to pay more attention to blood vessel features at different scales, improves segmentation performance sensitive to size changes, and achieves effective segmentation from small branches to larger main blood vessels. Attached Figure Description

[0013] Figure 1 This is a flowchart of the Fourier frequency domain enhanced blood vessel segmentation method of the present invention; Figure 2 This is a framework diagram of the overall training process of this invention; Figure 3 This is a diagram of the Fourier frequency domain enhanced blood vessel segmentation network structure of the present invention. Detailed Implementation

[0014] The technical solution of the present invention will now be described in detail through specific embodiments. Example

[0015] Please see Figure 1-2 This invention provides a Fourier frequency domain enhanced vessel segmentation method based on spatial topology awareness, comprising the following steps: S1. Acquire detailed CT image data and perform multi-dimensional preprocessing. The specific steps for acquiring detailed CT image data and performing multi-dimensional preprocessing are as follows: collect abdominal and lung CT images, and have clinical experts annotate vascular masks using Mimics to cover different types of blood vessels in the hepatic artery, vein, and pulmonary artery. Then, crop abnormal intensity values ​​of CT images, normalize the intensity using Z-score, and then enhance composite data, including using random Gaussian noise, random contrast enhancement, random mirror flip, random horizontal flip, and random rotation.

[0016] S2. Design the encoder and decoder, and construct the frequency domain enhancement network architecture; the specific steps for constructing the frequency domain enhancement network architecture and designing the encoder are as follows: First, a 3D image is input. Initial features are extracted using large kernel convolution, which serve as input to the encoder. DFT and IDFT are then used to transform the features between the spatial and frequency domains. Used to fill the frequency representation, in order to prevent the wrapping effect in subsequent convolution operations, the frequency domain representation is padded in advance. The zero-filling sequence can provide enough "space" for the linear convolution of the sequence to prevent the tail of a sequence from wrapping and interfering with the beginning of the convolution. formula express Image spatial representation of layers After normalization by the LN layer, the signal is then subjected to Discrete Fourier Transform (DFT). DFT can decompose complex signals into single frequency components with different amplitudes, thereby enabling operations such as image filtering. formula In the middle, frequency is represented With learnable weights Element-wise multiplication is performed between them, similar to convolution in the spatial domain. The result of the multiplication is cropped to remove the padding region and a frequency domain learnable bias term is added. formula In this context, the inverse discrete Fourier transform is used to represent the frequency. The data is transformed back to the spatial domain, and after being normalized through an LN layer, it is connected to a multilayer perceptron (MLP) and then added to the spatial data of the previous layer. This forms a residual, i.e., a jump connection. The formula represents the final spatial data. By using pooling operations to downsample the spatial representation, the Fourier transformer coding block effectively integrates traditional convolutional neural network operations with frequency domain processing techniques, which can fully utilize the advantages of both representations and improve network performance. In this embodiment, initial features of the 3D image are first extracted through large kernel convolution, laying the foundation for subsequent processing. Then, Discrete Fourier Transform (DFT) and Inverse Fourier Transform (IDFT) are used to convert the image between the spatial and frequency domains. Frequency domain filling operations are used to eliminate convolution wrapping effects, ensuring the accuracy of feature processing. In the frequency domain, filtering operations are performed through learnable weights and biases to specifically suppress noise and preserve vascular features. After inverse transformation back to the spatial domain, layer normalization (LN), multilayer perceptron (MLP), and residual connections are combined to enhance feature transfer and expression. Finally, pooling operations are used to complete downsampling. This approach effectively integrates the advantages of local spatial feature extraction and global frequency information processing, significantly improving the network's feature extraction and analysis performance for vascular images, and providing strong support for subsequent accurate segmentation.

[0017] S2. Design the encoder and decoder, and construct the frequency domain enhancement network architecture; the specific steps for constructing the frequency domain enhancement network architecture and designing the decoder are as follows: the specific formula for the Fourier fusion decoder is... The features derived from the encoder are denoted as E, and the features derived from the decoder are denoted as D, with superscripts... and These represent the layer depths, with the subscripts 'spatial' and 'freq' used to distinguish between spatial and frequency domain features after Fourier transform. Then, a Cropinner is applied to the shallow features of the encoder, and a Cropouter is applied to the deep features of the decoder, fusing low-frequency semantics with high-frequency details. Finally, the spatial domain features are transformed using the Inverse Discrete Fourier Transform (IDFT). The transformation is used as the input to the next layer of the decoder. In this embodiment, by distinguishing between the spatial and frequency domain representations of encoder and decoder features, Cropinner and Cropinner operations are used to process shallow high-frequency details and deep low-frequency semantic features respectively. The two are organically fused in the frequency domain, giving full play to the ability of frequency domain processing to capture global information. Subsequently, the fused frequency domain features are transformed back to the spatial domain through inverse discrete Fourier transform (IDFT) to form a feature representation that combines details and semantic information. This representation serves as the input for the next layer of the decoder, gradually restoring image resolution while preserving multi-scale vascular features. This design can effectively integrate the multi-scale features of the encoder with the layer-by-layer reconstruction capability of the decoder. The frequency domain fusion strategy enhances feature expression, especially in the preservation of details in small blood vessels and low-contrast regions, providing key support for achieving high-precision vascular segmentation.

[0018] S3. Establish a vascular anatomy-dependent model; the specific steps for establishing a vascular anatomy-dependent model are as follows: The FDESNet network was used to perform preliminary reasoning on CT images to generate vascular mask prediction results and locate the approximate area of ​​the blood vessels. Then, the labeled real vascular mask was expanded using a 5×5×5 three-dimensional expansion kernel to generate a DilatedLabel, which expanded the coverage of the vascular area and ensured that small branch voxels were included. Based on the expanded DilatedLabel, its minimum outer bounding box in three-dimensional space is calculated to determine the spatial range containing all expanded blood vessel voxels. Then, the original CT image is cropped using the above bounding box to remove irrelevant background areas and reduce noise interference. Finally, the cropped image size is dynamically adapted to the blood vessel distribution. For example, liver blood vessels can be cropped to 128×128×64 voxels. Extract the point set of blood vessels and organs, and extract the three-dimensional coordinates of all blood vessel voxels from the expanded blood vessel mask to form the blood vessel point set. , dimension ,in Point sets are also extracted from adjacent organs of blood vessels. Then, the farthest point sampling method is used to downsample the point cloud on the organ surface, retaining 8000 sample points to reduce the computational complexity of the subsequent calculations. The distance field map is normalized and cropped to reduce the interference of outliers. The Euclidean distance from each point in the blood vessel point set to all points in the organ point set is calculated and the minimum value is taken to generate the original distance field. Then the original distance values ​​are mapped to the interval [0,1] and a threshold is set to crop the distance values ​​that are out of range to the threshold boundary to reduce the interference of outliers. Finally, the standardized distance field map is merged with the cropped CT images and vascular mask prediction results as anatomically dependent contextual information and input into the subsequent topology learning module. By using the cropped small-size images and downsampled organ point sets, the computational complexity of the distance field is reduced, training efficiency is improved, and anatomically dependent information is integrated. In this embodiment, through the process of "preliminary prediction - region cropping - point set downsampling - distance encoding - information integration", FDESNet is first used to preliminarily locate blood vessels, and the mask is expanded and the image is cropped to reduce noise and speed up the process. Then, the point sets of blood vessels and organs are obtained by sampling the farthest point, the minimum Euclidean distance is calculated to generate a distance field map, and the data is standardized. Finally, the distance field map is fused with the image and the mask to provide contextual information containing anatomical spatial dependencies for subsequent topology learning, thereby reducing computational complexity and improving the accuracy of the model in locating and segmenting blood vessels.

[0019] S4. Multi-angle topology learning of blood vessel layers; The specific steps for multi-angle topology learning of blood vessel layers are as follows: The distance field map generated in the first part and the cropped CT image region are used as input. Then, the CT image is cropped by the outer rectangle of the expanding blood vessel mask to reduce background noise and accelerate training. FDESNet is used as the backbone network and combined with the Fourier Transformer encoder to generate multi-task features. The generated features are divided into three task branches, which are used for key point detection, centerline heatmap regression and adjacent connection prediction, respectively. For the vascular keypoint detection branch, FDESNet is used to generate a probability map of each point as a keypoint. The probability map represents the probability value of each point being predicted as a keypoint. Then, the weighted Hausdorff distance between the predicted point and the ground real point is calculated to calculate their similarity. For the centerline heatmap regression branch, a centerline heatmap unit is introduced, combined with a 3D adaptive Gaussian filter to generate a locally normalized heatmap. Defined as: and ,in These are the coordinates of point t on the center line. It means to pass through and the first boundary point on the cross section perpendicular to the centerline coordinates, yes The width of the core is used as the radius of the blood vessel, and finally the shape of the heat map is dynamically adjusted based on the diameter of the blood vessel to ensure the consistency of the skeleton line; For the neighbor connectivity prediction branch, the 26 nearest neighbors within a 3×3×3 cube region are selected as the target, given the blood vessel label. And its 26 nearest neighbors , will neighboring Index from 1 to 26, and pair each blood vessel. A binary tag encoded as a 1-channel label, if and If they are all blood vessel points, then they should be... Marked as a connection, corresponding Points in the channel It is marked as 1 if it is not connected, otherwise it is marked as 0, indicating no connection. Then training and prediction are performed. During the training phase, points are generated. and 26 nearest neighbors The prediction vectors of a total of 27 channels are generated, and then a channel probability map with connectivity likelihood is generated. In the prediction stage, the channel probability map is converted into a binary image as the final segmentation result. Pairs of voxels remain consistent with each other in terms of connectivity. In this embodiment, using the distance field map and cropped CT as input, FDESNet combined with Fourier Transform is used to extract features, and three branches are set up. Key point detection locates bifurcation points / endpoints through probability maps and WHD loss, preserving the tree structure. Centerline heatmap regression uses CHU units and adaptive Gaussian filtering to dynamically generate heatmaps based on vessel diameter, ensuring skeleton consistency. Adjacent connection prediction defines 27-channel connection relationships in a 3×3×3 neighborhood. The probability map is converted to a binary map to maintain the continuity of vessel topology. The three branches work together to capture the complete topological information of the vessel from key points to connectivity, effectively adapting to changes in vessel size and complex structures.

[0020] S5. Bottom-up attention interaction; the specific steps of bottom-up attention interaction are as follows: The key point detection probability map, centerline heatmap, and adjacent connection probability map output from the three branches are spliced ​​along the channel dimension. The resulting comprehensive feature map is subjected to global average pooling and fully connected operations. The importance weight of each channel is calculated. Then, based on the channel attention-weighted features, the attention weight of the spatial dimension is calculated to focus on the key regions of vascular topology. The channel attention and spatial attention are multiplied to obtain the final attention weight. The weighted composite feature map is then weighted and then the dimensionality of the weighted composite feature map is reduced by a 3D convolutional layer to generate a feature map that is suitable for the segmentation task. A segmentation mask is generated and post-processed. The feature map adapted to the segmentation task is input into the sigmoid activation function to generate a blood vessel segmentation probability map. Then, an adaptive threshold is used to convert the probability map P into a binary mask M. Based on the topological constraints of the adjacent connection prediction branches, morphological operations are performed on the binary mask to ensure the continuity of blood vessel connections. Finally, a three-dimensional Fourier transform is performed on the cropped image region to extract frequency domain features. The frequency domain information is then embedded into spatial features through a Fourier transformer encoder. The features embedded with frequency domain information are then weighted and fused with spatial attention features to enhance the segmentation accuracy of blood vessel details. In this embodiment, the key points, centerline, and connectivity features are spliced ​​together, and the weights are calculated using channel and spatial attention to focus on the key areas of the blood vessel. Then, a probability map is generated by dimensionality reduction. Thresholding and morphological operations are used to ensure the continuity of segmentation. Finally, frequency domain features are embedded and fused with spatial features to enhance the accuracy of blood vessel detail segmentation and achieve efficient interaction and integration of multi-scale topological information.

[0021] S6. Conduct multi-stage training and verify the experimental results; the specific steps for conducting multi-stage training and verifying the experimental results are as follows: The training process involves a phased, progressive approach. Using data labeled with an expanded vascular mask, the FDESNet network is trained to initially locate vascular regions. Simultaneously, a distance field map is generated to encode the spatial relationship between blood vessels and adjacent organs. Background noise interference is reduced by cropping CT images. Then, the distance field map and the cropped image are input, and the three branches of keypoint detection, centerline heatmap regression, and adjacent connectivity prediction are trained simultaneously. The Fourier Transformer encoder and attention mechanism are used to enhance the interaction of different topological features. Finally, the anatomy-dependent module and the topology learning module are connected in series, and bottom-up attention interaction is added to fine-tune the overall parameters to ensure the continuity of the model from coarse segmentation to fine topology learning. Public datasets such as 3Dircadb were divided into training, validation, and test sets in an 8:1:1 ratio to ensure that each subset contained vascular morphology of different individuals. Then, the intensity of CT images was normalized, outliers were cropped and scaled to a uniform range to eliminate the influence of equipment differences. The data was also augmented by random rotation, flipping, and scaling to enhance the model's adaptability to blood vessels at different angles. Gaussian noise was added and contrast was adjusted to simulate noise scenarios in clinical scans. The real vascular mask was expanded to increase the annotation range to cover small branches. The overlap rate between the predicted mask and the real label is calculated to measure the segmentation accuracy. The percentage of correctly segmented pixels is statistically analyzed to evaluate the model's ability to distinguish between the background and blood vessels. The proportion of correctly segmented real blood vessels is calculated to reflect the ability to capture small blood vessels. The proportion of the intersection of the predicted and real regions to the union is measured to supplement the evaluation dimension of the Dice metric. The benchmark methods are 3DU-Net, DV-Net and nnUNet. The same data preprocessing process and training rounds are used. The same hardware environment and loss function configuration are used to ensure the credibility of the comparison results. The results show that the method of the present invention achieves the best segmentation performance, with the highest values ​​of 72.34%, 99.47%, 73.12% and 56.79% in Dice, accuracy ACC, sensitivity Sen and Jaccard index, respectively, which confirms the superiority of the spatial topology-aware Fourier frequency domain enhanced blood vessel segmentation method. In this embodiment, a three-stage progressive training is adopted, and the generalization ability of the model is improved by combining data standardization and enhancement strategies. The model is quantitatively evaluated by multiple indicators such as Dice coefficient and accuracy, and is fairly compared with benchmark methods such as 3DU-Net to verify the superiority of the present invention in terms of blood vessel segmentation accuracy and robustness. The experimental results confirm the effectiveness of spatial topology perception and frequency domain enhancement technology with specific numerical values.

[0022] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness, characterized in that: Includes the following steps: S1. Acquire detailed CT image data and perform multi-dimensional preprocessing; S2. Design the encoder and decoder, and build a frequency domain enhancement network architecture; S3. Establish a vascular anatomy-dependent model; S4. Multi-angle topological learning of blood vessel layers; S5. Bottom-up attention interaction; S6. Conduct multi-stage training and verify the experimental results.

2. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S1, CT images of lung cancer patients and corresponding vascular areas such as hepatic arteries, veins, and pulmonary arteries are acquired; the acquired CT image data is cropped to remove abnormal intensity parts and normalized; and the normalized images are augmented with data including random Gaussian noise, random contrast enhancement, random mirror flip, random horizontal flip, and random rotation.

3. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S2, the specific steps for constructing the frequency domain enhancement network architecture and designing the encoder are as follows: (1) First, input a 3D image, extract initial features through large kernel convolution, and use these features as input to the encoder. Then, use DFT and IDFT to transform the features between the spatial and frequency domains: where Used to pad the frequency representation, in order to prevent the wrapping effect in subsequent convolution operations, the frequency domain representation is padded in advance. The zero-padding sequence can provide enough "space" for the linear convolution of the sequence to prevent the tail of a sequence from wrapping around and interfering with the beginning of the convolution. (2) express Image spatial representation of layers After normalization by the LN layer, the signal is then subjected to Discrete Fourier Transform (DFT). DFT can decompose complex signals into single frequency components with different amplitudes, thereby enabling operations such as image filtering. (3), formula In the middle, frequency is represented With learnable weights Element-wise multiplication is performed between them, similar to convolution in the spatial domain. The result of the multiplication is cropped to remove the padding region and a frequency domain learnable bias term is added. (4), formula In this context, the inverse discrete Fourier transform is used to represent the frequency. The data is transformed back to the spatial domain, and after being normalized through an LN layer, it is connected to a multilayer perceptron (MLP) and then added to the spatial data of the previous layer. This forms a residual, i.e., a jump connection. The formula represents the final spatial data. By using pooling operations to downsample the spatial representation, the Fourier transformer coding block effectively integrates traditional convolutional neural network operations with frequency domain processing techniques, which can fully utilize the advantages of both representations and improve network performance.

4. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S2, the specific steps for constructing the frequency domain enhancement network architecture and designing the decoder are as follows: The specific formula for the Fourier fusion decoder is... The features derived from the encoder are denoted as E, and the features derived from the decoder are denoted as D, with superscripts... and These represent the layer depths, with the subscripts 'spatial' and 'freq' used to distinguish between spatial and frequency domain features after Fourier transform. Then, a Cropinner is applied to the shallow features of the encoder, and a Cropouter is applied to the deep features of the decoder, fusing low-frequency semantics with high-frequency details. Finally, the spatial domain features are transformed using the Inverse Discrete Fourier Transform (IDFT). The transformation is used as the input to the next layer of the decoder.

5. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S3, the specific steps for establishing the vascular anatomy-dependent model are as follows: (1) Use the constructed FDESNet network to perform preliminary reasoning on CT images, generate vascular mask prediction results, locate the approximate area of ​​the blood vessels, and then use a 5×5×5 three-dimensional expansion kernel to expand the labeled real vascular mask to generate DilatedLabel, expand the coverage of the vascular area, and ensure that small branch voxels are included. (2) Based on the expanded DilatedLabel, calculate its minimum outer rectangle in three-dimensional space, determine the spatial range containing all expanded blood vessel voxels, and then use the above rectangle to crop the original CT image, remove irrelevant background areas, reduce noise interference, and finally the cropped image size dynamically adapts to the blood vessel distribution. For example, liver blood vessels can be cropped to 128×128×64 voxels. (3) Extract the blood vessel and organ point set. Extract the three-dimensional coordinates of all blood vessel voxels from the expanded blood vessel mask to form the blood vessel point set. , dimension ,in Point sets are also extracted from adjacent organs of blood vessels. Then, the farthest point sampling method is used to downsample the point cloud on the organ surface, retaining 8000 sample points to reduce the computational complexity of the subsequent calculations. (4) Normalize and crop the distance field map to reduce the interference of outliers. Calculate the Euclidean distance from each point in the blood vessel point set to all points in the organ point set, and take the minimum value to generate the original distance field. Then map the original distance values ​​to the [0,1] interval and set a threshold to crop the distance values ​​that are out of range to the threshold boundary to reduce the interference of outliers. (5) Finally, the standardized distance field map is combined with the cropped CT image and the vascular mask prediction results as anatomical dependent context information and input into the subsequent topology learning module. By using the cropped small-size image and downsampled organ point set, the computational complexity of the distance field is reduced, the training efficiency is improved, and the anatomical dependent information is integrated.

6. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S4, the specific steps for multi-angle topological learning of blood vessel layers are as follows: (1) The distance field map generated in the first part and the cropped CT image region are used as input. Then, the CT image is cropped by the outer rectangular box of the expanding blood vessel mask to reduce background noise and accelerate training. FDESNet is used as the backbone network and Fourier Transformer encoder is combined to generate multi-task features. The generated features are divided into three task branches, which are used for key point detection, centerline heat map regression and adjacent connection prediction, respectively. (2) For the blood vessel key point detection branch, a probability map of each point as a key point is generated by FDESNet. The probability map represents the probability value of each point being predicted as a key point. Then, the weighted Hausdorff distance between the predicted point and the ground real point is calculated to calculate their similarity. (3) For the centerline heatmap regression branch, a centerline heatmap unit is introduced, and a local normalized heatmap is generated by combining it with a 3D adaptive Gaussian filter. Defined as: and ,in These are the coordinates of point t on the center line. It means to pass through and the first boundary point on the cross section perpendicular to the centerline coordinates, yes The width of the core is used as the radius of the blood vessel, and finally, the shape of the heat map is dynamically adjusted based on the blood vessel diameter to ensure the consistency of the skeleton lines. (4) For the adjacent connection prediction branch, taking the 26 nearest neighbors in the 3×3×3 cube region as the object, given the blood vessel label And its 26 nearest neighbors , will the field Index from 1 to 26, and pair each blood vessel. A binary tag encoded as a 1-channel label, if and If they are all blood vessel points, then they should be... Marked as a connection, corresponding Points in the channel It is marked as 1 if it is not connected, otherwise it is marked as 0, indicating no connection. Then training and prediction are performed. During the training phase, points are generated. and 26 nearest neighbors The prediction vectors of a total of 27 channels are generated, and then a channel probability map with connectivity likelihood is generated. In the prediction stage, the channel probability map is converted into a binary image as the final segmentation result, and pairs of voxels remain consistent with each other in terms of connectivity.

7. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S5, the specific steps for bottom-up attention interaction are as follows: (1) The key point detection probability map, centerline heatmap and adjacent connection probability map output by the three branches are spliced ​​along the channel dimension. The generated comprehensive feature map is subjected to global average pooling and full connection operation. The importance weight of each channel is calculated. Then, based on the channel attention weighted features, the attention weight of the spatial dimension is calculated to focus on the key regions of vascular topology. (2) Multiply the channel attention and spatial attention to obtain the final attention weight, weight the spliced ​​comprehensive feature map, and then reduce the dimensionality of the weighted comprehensive feature map through a three-dimensional convolutional layer to generate a feature map adapted to the segmentation task. (3) Generate a segmentation mask and perform post-processing. Input the feature map adapted to the segmentation task into the sigmoid activation function to generate a blood vessel segmentation probability map. Then, use an adaptive threshold to convert the probability map P into a binary mask M. Based on the topological constraints of the adjacent connection prediction branches, perform morphological operations on the binary mask to ensure the continuity of blood vessel connections. (4) Finally, a three-dimensional Fourier transform is performed on the cropped image region to extract frequency domain features. The frequency domain information is embedded into the spatial features through a Fourier transform encoder. The features embedded with frequency domain information are weighted and fused with spatial attention features to enhance the segmentation accuracy of blood vessel details.

8. The Fourier frequency domain enhanced blood vessel segmentation method based on spatial topology awareness according to claim 1, characterized in that: In step S6, the specific steps for performing multi-stage training and verifying experimental results are as follows: (1) Perform phased progressive training. Use the expanded blood vessel mask annotation data to train the FDESNet network to initially locate the blood vessel region. At the same time, generate a distance field map to encode the spatial relationship between blood vessels and adjacent organs. Reduce background noise interference by cropping CT images. Then input the distance field map and the cropped image. Simultaneously train the three branches of key point detection, centerline heat map regression, and adjacent connection prediction. Use the Fourier Transformer encoder and attention mechanism to enhance the interaction of different topological features. Finally, connect the anatomical dependency module and the topology learning module, add bottom-up attention interaction, and fine-tune the overall parameters to ensure the continuity of the model from coarse segmentation to fine topology learning. (2) Divide the public datasets such as 3Dircadb into training set, validation set and test set in an 8:1:1 ratio to ensure that each subset contains the vascular morphology of different individuals. Then normalize the intensity of the CT images, crop outliers and scale them to a uniform range to eliminate the influence of equipment differences. Enhance the data by randomly rotating, flipping and scaling to enhance the model’s adaptability to blood vessels at different angles. Add Gaussian noise and adjust the contrast to simulate the noise scene in clinical scanning. Expand the real mask of blood vessels to expand the annotation range to cover small branches. (3) Calculate the overlap rate between the predicted mask and the real label, measure the segmentation accuracy, count the proportion of correctly segmented pixels, evaluate the model's ability to distinguish between the background and blood vessels, calculate the proportion of correctly segmented real blood vessels to reflect the ability to capture small blood vessels, measure the proportion of the intersection of the predicted and real regions to the union, supplement the evaluation dimensions of the Dice index, the benchmark methods are 3DU-Net, DV-Net and nnUNet, the same data preprocessing process and training rounds are used, the same hardware environment and loss function configuration are used uniformly to ensure the credibility of the comparison results; (4) The results show that the method of the present invention has achieved the best segmentation performance, with the highest values ​​of 72.34%, 99.47%, 73.12% and 56.79% in Dice, accuracy ACC, sensitivity Sen and Jaccard index respectively, which confirms the superiority of the spatial topology-aware Fourier frequency domain enhanced blood vessel segmentation method.

Images