Cerebral vascular endothelial segmentation method based on double-coordinate attention mechanism

Abstract

The application discloses a brain blood vessel endothelium segmentation method based on a double-coordinate attention mechanism, relates to the technical field of image processing, and comprises the following steps: in view of the problems of blurred endothelium boundary, low contrast and significant geometric deformation in optical coherence tomography images, a double-path feature extraction framework fusing Cartesian and polar coordinate representation is proposed; a coordinate system adaptive conversion is realized through a differentiable polar coordinate transformation module, and a double-path coordinate attention module is designed to extract one-dimensional direction descriptors along the horizontal, vertical, radial and angular directions; the optimized feature map is input into a subsequent segmentation network, four-way attention weight maps are generated through cross-domain fusion, feature re-labeling is realized, and finally, brain blood vessel endothelium image segmentation is realized. Through the construction of the Cartesian-polar double-path feature extraction framework and the cross-domain collaborative double-path coordinate attention mechanism, the blurred brain blood vessel endothelium boundary is modeled, and the segmentation precision of the endothelium structure in the OCT image is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for cerebral vascular endothelial segmentation based on a dual-coordinate attention mechanism. Background Technology

[0002] Cerebrovascular diseases are one of the leading causes of death and disability worldwide, with ischemic stroke accounting for 60%-80% of these cases. Accurate assessment of pathological changes such as vascular stenosis is crucial for early diagnosis and treatment. Optical coherence tomography (OCT), a high-resolution imaging technique, can clearly visualize the layered structure of blood vessel walls with micron-level precision, enabling detailed observation of the cerebral vascular endothelium. However, the endothelium in OCT images is typically thin, with blurred boundaries, and susceptible to speckle noise; manual segmentation is time-consuming, labor-intensive, and subject to subjective variations.

[0003] Currently, deep learning-based medical image segmentation methods, such as U-Net and its variants, have achieved success in various tasks. However, directly applying these mainstream models to OCT brain vascular endothelial segmentation still faces challenges: First, most network designs are based on Cartesian coordinate space, which makes it difficult to effectively fit the inherent radial symmetry and annular features of vascular structures, limiting the ability to model the continuity and directionality of the vessel wall; second, existing attention mechanisms mostly focus on feature enhancement in channels or single spatial domains (such as Cartesian space), lacking explicit modeling of radial and angular features in polar coordinates, and failing to fully exploit the complementary geometric information in the two coordinate systems. Although some studies have attempted to introduce polar coordinate transformation or channel attention to improve performance, how to synergistically utilize dual-domain features and conduct effective cross-coordinate attention interaction remains an urgent problem to be solved. Therefore, a new technical solution is urgently needed that can deeply integrate the representational advantages of Cartesian and polar coordinates, and achieve accurate and robust segmentation of brain vascular endothelial cells in OCT images through a more intelligent attention mechanism, thereby providing reliable technical support for clinical evaluation. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism. By constructing a Cartesian-polar coordinate dual-path feature extraction framework and a cross-domain collaborative dual-path coordinate attention mechanism, robust modeling is achieved for challenges such as blurred boundaries, low contrast, and geometric deformation of brain vascular endothelial structures, significantly improving the segmentation accuracy and boundary localization accuracy of endothelial structures in OCT images.

[0005] A brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism includes:

[0006] S1, acquire cerebral vascular optical coherence tomography (OCT) images, perform center alignment, intensity normalization, and data augmentation on the cerebral vascular OCT images to obtain preprocessed images;

[0007] S2, the preprocessed image is synchronously input to the Cartesian coordinate branch and the polar coordinate branch; the Cartesian coordinate branch extracts features from the preprocessed image through convolution operation and outputs the first feature map; the polar coordinate branch transforms the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a differentiable polar coordinate transformation module, and then extracts features from the polar coordinate image through convolution operation and outputs the second feature map;

[0008] S3, input the first feature map and the second feature map together into the dual-path coordinate attention module, extract one-dimensional directional feature descriptors along the horizontal direction of the Cartesian coordinate system, the vertical direction of the Cartesian coordinate system, the radial direction of the polar coordinate system, and the angular direction of the polar coordinate system, respectively, and perform cross-coordinate domain feature fusion to generate attention weight maps corresponding to the four directions;

[0009] S4. Based on the attention weight map, perform element-wise multiplication of the first feature map and the second feature map for attention recalibration to obtain the enhanced first output feature map and the second output feature map.

[0010] S5 inputs the enhanced first and second output feature maps into the segmentation network for training and inference, and finally outputs the segmentation result of cerebral vascular endothelium.

[0011] Furthermore, in S2, the polar coordinate branch transforms the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a differentiable polar coordinate transformation module. The specific transformation method is as follows:

[0012] The center position and maximum radius of the preprocessed image in the Cartesian coordinate system are determined. In the polar coordinate domain, the center position is used as the pole, and uniform sampling is performed radially within a range not exceeding the maximum radius, and uniform sampling is also performed angularly to construct a polar coordinate sampling grid. Each sampling point in the polar coordinate sampling grid is mapped back to the corresponding coordinate position in the Cartesian coordinate system according to its radial distance and angle. The corresponding coordinate position in the preprocessed image is sampled through bilinear interpolation to obtain the polar coordinate domain feature map.

[0013] Furthermore, in S3, one-dimensional directional feature descriptors are extracted, specifically including:

[0014] Global average pooling is performed on the first feature map along its horizontal and vertical directions to obtain row direction descriptors and column direction descriptors.

[0015] Global average pooling is performed on the second feature map along its radial and angular directions to obtain radial and angular descriptors, respectively.

[0016] Furthermore, in S3, attention weight maps corresponding to the four directions are generated, specifically including:

[0017] The row direction descriptor, column direction descriptor and radial descriptor are concatenated, processed by the first lightweight bottleneck network to generate the first fusion feature, which is then split into row direction attention map and column direction attention map;

[0018] The row direction descriptors, column direction descriptors, and angular direction descriptors are concatenated, processed by the second lightweight bottleneck network, and a second fusion feature is generated, which is then split into radial attention maps and angular attention maps.

[0019] Furthermore, attention recalibration is performed on the first and second feature maps based on the attention weight map through element-wise multiplication. The calculation formula is as follows:

[0020] ;

[0021] ;

[0022] in, and These are the enhanced first and second output feature maps, respectively. and These are the first feature map and the second feature map, respectively. , , and These are row-direction attention maps, column-direction attention maps, radial attention maps, and angular attention maps, respectively.

[0023] Furthermore, in S3, the dual-path coordinate attention module is deployed in the encoder stage of the U-Net segmentation network and inserted after the skip connections.

[0024] Furthermore, the segmentation results are quantitatively evaluated using regional overlap and boundary distance evaluation metrics. The formula for calculating regional overlap is as follows:

[0025] ;

[0026] ;

[0027] Where Dice represents the overlap between the prediction and the ground truth annotation; P is the set of predicted segmented pixels, and G is the set of ground truth labeled pixels; Represents the set of boundary points of the predicted segmentation result; Represents the set of boundary points that are actually labeled; express One of the boundary pixels; express One of the boundary pixels; This represents the Euclidean distance between pixels; This means sorting the nearest distances of all boundary points and taking the distance value corresponding to the 95th percentile. This indicates the degree of alignment between the predicted boundary and the true boundary.

[0028] The present invention adopts the above technical solution and has the following beneficial effects:

[0029] (1) By introducing a differentiable polar coordinate transformation module and a dual-path coordinate attention mechanism, the present invention simultaneously models the structural priors in the horizontal and vertical directions of the Cartesian coordinate system and the radial and angular directions of the polar coordinate system, effectively capturing the annular geometric characteristics and local details of the vascular endothelium.

[0030] (2) This invention extracts features through dual-coordinate branch collaboration and generates a four-way attention weight map through cross-domain fusion, making the network more adaptable to common interferences in OCT images such as speckle noise, low contrast and center shift.

[0031] (3) The present invention combines the dual evaluation system of regional overlap and boundary distance, which not only reflects the overall consistency of the segmentation results, but also focuses on the accuracy of the location of the endothelial boundary, which is of clinical concern, providing an objective and quantifiable evaluation basis for algorithm verification and clinical translation. Attached Figure Description

[0032] Figure 1 This invention provides a method for cerebral vascular endothelial segmentation based on a dual-coordinate attention mechanism.

[0033] Figure 2 This is a schematic diagram of the dual-path coordinate attention module according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the integration of the dual-path coordinate attention module in U-Net according to an embodiment of the present invention;

[0035] Figure 4 This is a qualitative comparison diagram of the segmentation results of representative samples in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0037] like Figure 1 As shown, the present invention provides a brain vascular endothelial segmentation method based on a dual-coordinate attention mechanism, comprising:

[0038] S1. Acquire cerebral vascular optical coherence tomography (OCT) images, perform center alignment, intensity normalization, and data augmentation on the cerebral vascular OCT images to obtain preprocessed images.

[0039] Specifically, during the preprocessing process, to ensure the robustness and generalization ability of the proposed segmentation model, all original OCT images underwent a unified preprocessing procedure before training. First, to ensure the geometric alignment accuracy of subsequent polar coordinate transformations, the images were manually verified and adjusted to ensure that the vessel centerline was located as close to the center of the image as possible. Subsequently, the z-score normalization method was used to normalize the image intensity, ensuring that the pixel intensity distribution of each image had zero mean and unit variance. To suppress the impact of extreme intensity values ​​on model training, the normalized pixel intensity was truncated to the interval [−3σ, +3σ] and further linearly mapped to the range [0, 1] to reduce the interference of intensity differences caused by factors such as tissue scattering characteristics, differences in imaging parameters, and changes in imaging depth on model performance.

[0040] S2, the preprocessed image is synchronously input to the Cartesian coordinate branch and the polar coordinate branch; the Cartesian coordinate branch extracts features from the preprocessed image through convolution operation and outputs the first feature map; the polar coordinate branch transforms the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a differentiable polar coordinate transformation module, and then extracts features from the polar coordinate image through convolution operation and outputs the second feature map.

[0041] Specifically, the polar coordinate branch transforms the preprocessed image from Cartesian coordinates to polar coordinates using a differentiable polar coordinate transformation module. The specific transformation method is as follows:

[0042] The center position and maximum radius of the preprocessed image in the Cartesian coordinate system are determined. In the polar coordinate domain, the center position is used as the pole, and uniform sampling is performed radially within a range not exceeding the maximum radius, and uniform sampling is also performed angularly to construct a polar coordinate sampling grid. Each sampling point in the polar coordinate sampling grid is mapped back to the corresponding coordinate position in the Cartesian coordinate system according to its radial distance and angle. The corresponding coordinate position in the preprocessed image is sampled through bilinear interpolation to obtain the polar coordinate domain feature map.

[0043] Specifically, many vascular structures and other biological targets exhibit approximate centrosymmetry, and their image intensity distribution typically displays a significant radial regularity around a geometric center—that is, from the center outwards, the intensity change mainly depends on the distance from the center; while when rotating around the center, the intensity shows periodic or orientation-dependent changes with the angle. This inherent geometric characteristic suggests that, compared to the translation-invariant modeling along the horizontal and vertical directions in the traditional Cartesian coordinate system, using polar coordinates (with the center as the origin, and using radius and angle to describe position) better reflects the physical nature of annular structures such as blood vessels. In polar coordinates, convolution operations can more naturally capture the radial continuity of the blood vessel wall (such as the gradual change of the endothelium from the inside out), angular consistency (such as the closure of a complete annular boundary), and rotation invariance, thereby improving the model's robustness to disturbances such as scale scaling, slight rotation, and local deformation. To this end, this paper designs a differentiable polar coordinate transformation module, implemented based on the standard sampling operator in the deep learning framework. This module supports bidirectional mapping: on the one hand, it can uniformly sample and resample the original Cartesian coordinate image into a polar coordinate feature map according to a preset radius range and angular resolution; on the other hand, it can also reverse-map the polar coordinate feature map back to a Cartesian coordinate output of a specified size. The entire process uses bilinear interpolation, maintaining gradient connectivity throughout, and can be seamlessly embedded into end-to-end training processes to ensure joint optimization of network parameters.

[0044] S3. Input the first feature map and the second feature map together into the dual-path coordinate attention module, extract one-dimensional directional feature descriptors along the horizontal direction of the Cartesian coordinate system, the vertical direction of the Cartesian coordinate system, the radial direction of the polar coordinate system, and the angular direction of the polar coordinate system, respectively, and perform cross-coordinate domain feature fusion to generate attention weight maps corresponding to the four directions.

[0045] Specifically, in S3, one-dimensional directional feature descriptors are extracted, including:

[0046] Global average pooling is performed on the first feature map along its horizontal and vertical directions to obtain row direction descriptors and column direction descriptors.

[0047] Global average pooling is performed on the second feature map along its radial and angular directions to obtain radial and angular descriptors, respectively.

[0048] Specifically, generating attention weight maps corresponding to the four directions includes:

[0049] The row direction descriptor, column direction descriptor and radial descriptor are concatenated, processed by the first lightweight bottleneck network to generate the first fusion feature, which is then split into row direction attention map and column direction attention map;

[0050] The row direction descriptors, column direction descriptors, and angular direction descriptors are concatenated, processed by the second lightweight bottleneck network, and a second fusion feature is generated, which is then split into radial attention maps and angular attention maps.

[0051] Specifically, attention recalibration is performed on the first and second feature maps through element-wise multiplication based on the attention weight map. The calculation formula is as follows:

[0052] ;

[0053] ;

[0054] in, and These are the enhanced first and second output feature maps, respectively. and These are the first feature map and the second feature map, respectively. , , and These are row-direction attention maps, column-direction attention maps, radial attention maps, and angular attention maps, respectively.

[0055] Specifically, the dual-path coordinate attention module is deployed in the encoder stage of the U-Net segmentation network and inserted after the skip connections.

[0056] Specifically, such as Figure 2 This diagram illustrates the overall structure and data flow of the Dual-Path Coordinate Attention (DCA) module: It clearly presents the module's dual-input, dual-path processing, and cross-domain fusion mechanism—the left input is a Cartesian coordinate branch feature map, and the right input is a polar coordinate branch feature map obtained through differentiable polar coordinate transformation. The diagram explicitly marks the four one-dimensional feature extraction paths (horizontal, vertical, radial, and angular), as well as two types of cross-coordinate concatenation strategies (row / column + radial, row / column + angular). Subsequently, a lightweight bottleneck network generates intermediate fusion features, which are then split and activated to output four-way directional attention maps. Finally, these attention maps are applied to the original feature maps of their corresponding branches to complete adaptive recalibration. The entire diagram intuitively embodies the core design philosophy of the DCA module: "dual-coordinate representation—direction awareness—cross-domain fusion—synergistic enhancement," highlighting its ability to accurately model the multi-scale, multi-directional geometric structure of vascular endothelium without significantly increasing the number of parameters.

[0057] Specifically, after obtaining one-dimensional directional feature descriptors, the DCA module concatenates these descriptors along the channel dimension to achieve cross-coordinate domain feature interaction. First, the row, column, and radial descriptors are concatenated to form a joint feature integrating Cartesian coordinate horizontal / vertical information and polar coordinate radial information. Then, the row, column, and angular descriptors are concatenated to generate another set of joint features containing Cartesian coordinate horizontal / vertical information and polar coordinate angular information. These two types of features are then input into a lightweight bottleneck network to compress dimensions, preserving key feature representations without significantly increasing computational complexity. The outputs of this process represent the results of horizontal / vertical plus radial feature fusion and horizontal / vertical plus angular feature fusion, respectively. Next, the output features are split along the channel dimension and processed to generate four attention maps, corresponding to the row, column, radial, and angular directions, respectively, to enhance key regions of the original feature map. Finally, using element-wise multiplication, these attention maps are used to adaptively recalibrate the original feature map, thereby highlighting important structural information in multiple directions and suppressing irrelevant background noise, improving the accuracy and robustness of feature representation. This mechanism enables the model to focus on key information more effectively in different spatial patterns.

[0058] Specifically, such as Figure 3 The diagram illustrates the integration of the Dual-Path Coordinate Attention (DCA) module into the standard U-Net network architecture. This module is deployed after skip connections in each upsampling path, specifically inserted after concatenating the high-resolution features from the encoder with the corresponding scale features from the decoder, and before subsequent convolutional processing. This design allows multi-scale features from the encoder to undergo simultaneous joint attention recalibration in both Cartesian (horizontal and vertical) and polar (radial and angular) coordinates. Lower-level features are enhanced with local details such as edges and textures, while higher-level features focus more on the overall annular structure, curvature, and radial symmetry of the blood vessels. This location selection balances spatial detail preservation with anatomical prior guidance, achieving coherent and accurate segmentation of slender, blurred, or irregular cerebral vascular endothelial structures in OCT images without significantly increasing computational burden.

[0059] S4. Based on the attention weight map, perform element-wise multiplication of the first feature map and the second feature map for attention recalibration to obtain the enhanced first output feature map and the second output feature map.

[0060] S5 inputs the enhanced first and second output feature maps into the segmentation network for training and inference, and finally outputs the segmentation result of cerebral vascular endothelium.

[0061] Specifically, the proposed Dual-Path Coordinate Attention (DCA) aims to simultaneously utilize the complementary structural information contained in the Cartesian and polar coordinate systems. It can capture global contextual information along the horizontal and vertical directions of the Cartesian coordinate system and the radial and angular directions of the polar coordinate system, respectively, and generate a comprehensive and direction-aware spatial attention representation through cross-coordinate domain feature fusion. In terms of design, DCA first receives feature maps x1 from the Cartesian coordinate branch and x2 from the polar coordinate branch as input. The entire processing flow consists of three stages: the first stage is one-dimensional directional feature extraction, extracting feature descriptors from the four main directions of horizontal, vertical, radial, and angular to enhance the network's ability to perceive spatial location information; the second stage is to fuse the above directional features across different coordinate domains through a lightweight bottleneck network, generating intermediate feature representations and further decomposing them into multiple directional attention weight maps; the final stage is to recalibrate the original features using the generated directional attention maps, thereby highlighting key structural regions and suppressing irrelevant background information. This design allows the network to fully leverage the advantages of Cartesian and polar coordinate representations while focusing on horizontal, vertical, radial, and angular patterns, achieving fine modeling of complex structures such as vascular structures. In the one-dimensional directional feature extraction process, the DCA module employs adaptive average pooling to extract one-dimensional feature descriptors reflecting directional information from two coordinate representations. For the Cartesian coordinate branch, the row direction descriptor xh and column direction descriptor xw are obtained by globally averaging the feature map x1 along the height and width directions. These two one-dimensional vectors encode global contextual information across rows and columns. For the polar coordinate branch, the radial descriptor xr and angular descriptor xθ are obtained by globally averaging along the radial and angular directions. xr reflects the intensity distribution characteristics from the center outwards, while xθ describes the angular distribution characteristics around the center. In this way, the spatial dimension is compressed into one-dimensional vectors of lengths H, W, R, and Θ, reducing computational overhead while preserving representative statistical information in each principal direction.

[0062] Specifically, this study collected 2000 cross-sectional OCT images of cerebral blood vessels. Each image had a spatial resolution of 451 × 394 pixels and was stored in 8-bit grayscale format. All images were acquired from intracranial arterial regions, allowing for high-fidelity representation of the vessel wall structure, particularly suitable for imaging analysis of areas prone to stenosis. Two experienced neuroradiologists used ITK-SNAP software to perform detailed manual annotation of the cerebral vascular endothelial regions. All annotation results were reviewed by a third senior expert to ensure consistency and reliability of the actual labels. Disagreements among the annotators were resolved through joint discussion.

[0063] To evaluate the effectiveness of the proposed method, the dataset was randomly divided into training and testing sets in an 8:2 ratio, ensuring that images of the same patient did not appear in different subsets simultaneously, thus effectively avoiding data leakage. This dataset is used for the development and evaluation of a vascular endothelial segmentation algorithm in brain OCT images.

[0064] Specifically, in this embodiment, to further enhance the model's generalization ability and noise resistance, a real-time data augmentation strategy is introduced during training, including random horizontal flipping, random vertical flipping, random rotation, and Gaussian noise perturbation. While maintaining the vascular structure information without fundamentally altering it, these augmentation methods effectively improve the model's adaptability to different deformation and noise conditions. Furthermore, this paper employs the differentiable polar coordinate transformation module proposed in the method section to map OCT images in the Cartesian coordinate domain to polar coordinate space, thereby explicitly modeling the radial features of the vascular structure. Finally, the Cartesian and polar coordinate representations of each image are fed together as input into the dual-branch network structure. This dual-view feature representation strategy effectively integrates complementary information from the two coordinate domains, providing multi-angle spatial information support for the dual-path coordinate attention module, thereby further improving the segmentation accuracy and anatomical consistency of the vascular endothelial region.

[0065] Specifically, to comprehensively and objectively evaluate the performance of the proposed method in the cerebral vascular endothelial segmentation task, this paper adopts six core evaluation metrics recognized in the field of medical image segmentation: Dice similarity coefficient, 95th percentile Hausdorff distance, precision, sensitivity, crossover ratio (CRR), and volume similarity. These metrics synergistically characterize segmentation quality from different perspectives, forming a multi-dimensional and complementary evaluation system. The Dice similarity coefficient measures the overall overlap between the predicted segmented region and the ground truth labeled region; a higher value indicates greater consistency in spatial coverage. Its value ranges from 0 to 1, with 1 representing complete overlap. The 95th percentile Hausdorff distance assesses the spatial alignment accuracy between the predicted boundary and the ground truth boundary, reflecting the maximum deviation level between the vast majority of boundary points. Compared to the traditional Hausdorff distance, this metric is less sensitive to outliers and is more robust and clinically interpretable; a lower value indicates more accurate boundary localization. Precision indicates the proportion of pixels predicted as endothelium by the model that actually belong to the endothelium, reflecting the algorithm's ability to avoid false positives. Sensitivity indicates the proportion of real endothelium pixels that are correctly identified, reflecting the algorithm's ability to detect weak or ambiguous boundaries. Intersection over Union (IoU) further quantifies the overlap between the predicted and real regions at the pixel level; a higher value indicates a tighter region match. Volume similarity assesses the consistency between the predicted results and the real annotations on an overall volume scale, reflecting whether the model produces systematic oversegmentation or undersegmentation; the closer the value is to 1, the more reliable the volume estimation. In summary, these six indicators provide a systematic, quantitative, and interpretable comprehensive evaluation of the cerebral vascular endothelium segmentation results from the perspectives of region coverage, boundary accuracy, classification reliability, spatial consistency, and anatomical rationality.

[0066] Table 1 shows the comparative experimental results;

[0067] method Dice HD95 Precision Sensitivity IoU VS U-Net 0.7701 10.4035 0.8196 0.7483 0.6869 0.8158 U-Net++ 0.8305 11.4765 0.9548 0.7980 0.7582 0.8497 AttentionUNet 0.8044 12.9114 0.9085 0.7883 0.7375 0.8305 DeeplabV2 0.7208 14.4364 0.7674 0.7318 0.6238 0.7619 DeeplabV3 0.7803 19.4165 0.6895 0.8985 0.6397 0.8684 Swin-Unet 0.8207 10.1986 0.8799 0.8479 0.7413 0.8564 This article's method 0.8707 6.6880 0.8698 0.9053 0.7844 0.9025

[0068] Specifically, as shown in Table 1, to verify the effectiveness of the proposed Dual Coordinate Attention (DCA) segmentation framework, this paper conducts a systematic comparative experiment with six mainstream medical image segmentation models (U-Net, U-Net++, Attention U-Net, DeeplabV2, DeeplabV3, and Swin-Unet) under a unified experimental setting. All models are implemented based on PyTorch, using the same dataset partitioning, preprocessing procedures, batch size, Adam optimizer, learning rate scheduling strategy, and Dice loss function for training and testing. The average results of six metrics—Dice, HD95, Precision, Sensitivity, IoU, and VS—are uniformly reported on the test set. As shown in Table 1, the proposed method significantly outperforms all comparative models in terms of leading performance metrics such as Dice 0.8707, HD95 6.6880, and VS 0.9025. In particular, it has outstanding advantages in boundary localization accuracy (HD95 reduces by 3.5106 compared to the second-best model Swin-Unet), segmentation consistency (VS improves by 4.61% compared to the second-best model DeeplabV3), and overall segmentation quality (IoU reaches 0.7844). This fully demonstrates the effectiveness and advancement of dual-coordinate feature modeling and cross-domain attention mechanism in the fine segmentation task of cerebral vascular endothelial cells.

[0069] Specifically, to verify the synergistic effect between the dual-path network structure and the Channel Coordinate Attention (DCA) mechanism, this paper constructs seven model variants for ablation experiments based on two dimensions: "number of paths" and "attention deployment stage." These include: a single-path U-Net baseline model, a single-network attention enhancement model (CroodAttUNet), a dual-U-Net structure without attention interactions (Dual-PolarNet), a model incorporating DCA in all decoder stages (DCA-Decoder), a model incorporating DCA only in the final decoder stage (DCA-Decoder-Terminal), a model incorporating DCA only in the encoder stage (DCA-Encoder), and a model incorporating DCA in both the encoder and decoder stages (Full DCA). All model variants are trained and tested on the same dataset, with the same training configuration and evaluation criteria, focusing on key performance indicators in medical image segmentation such as Dice, HD95, and Precision.

[0070] Table 2 Comparison of ablation test results;

[0071] method Dice HD95 Precision Sensitivity IoU VS Baseline 0.7701 10.4035 0.8196 0.7483 0.6869 0.8158 CroodAttUNet 0.7574 9.4661 0.7305 0.8226 0.6754 0.7932 Dual-PolarNet 0.8554 9.6622 0.8139 0.9336 0.7641 0.8955 DCA-Decoder 0.8200 10.9822 0.8634 0.8448 0.7269 0.8725 DCA-Decoder-Terminal 0.8502 7.9957 0.8815 0.8716 0.7649 0.8900 DCA-Encoder 0.8707 6.6880 0.8698 0.9053 0.7844 0.9025 Full DCA 0.7441 12.6649 0.9177 0.7102 0.6352 0.7582

[0072] As shown in Table 2 (the best results are marked in bold), the ablation experiment results systematically revealed the influence of different model structures on the segmentation performance of cerebral vascular endothelium: the baseline (single-path U-Net) achieved Dice 0.7701 and HD95 10.4035, representing the basic segmentation ability of this task; the performance of CrooodAttUNet (Dice 0.7574), which introduces coordinate attention, actually decreased slightly, indicating that simply stacking attention under a single path is not effective; while Dual-PolarNet, which does not use attention but adopts a Cartesian-polar coordinate dual-path structure, significantly improved Dice to 0.8554 (about 11.08% higher than the baseline) and HD95 to 9.6622, confirming that the dual-coordinate representation itself contains strong geometric priors, which is the key basis for the performance leap. Further analysis of attention deployment locations revealed that full decoder deployment (DCA-Decoder) caused HD95 to rise to 10.9822, resulting in unstable performance. Deploying only at the decoder terminal (DCA-Decoder-Terminal) balanced Dice (0.8502) and boundary accuracy (HD95 = 7.9957). Deploying DCA only at the encoder stage (DCA-Encoder) achieved the best overall performance (Dice = 0.8707, HD95 = 6.6880, Sensitivity = 0.9053, IoU = 0.7844, VS = 0.9025), indicating that early fusion of dual-domain features in the encoder is most beneficial for structural prior modeling. Conversely, full-stage deployment (FullDCA) caused Dice to plummet to 0.7441 and HD95 to rise to 12.6649, verifying that attention interaction requires refined design, and excessive deployment will lead to feature convergence and degradation of discriminative power.

[0073] Specifically, such as Figure 4The image shown is a qualitative comparison of the segmentation results of representative samples in this embodiment of the invention. From left to right, the results are: original OCT image, UNet, UNetPlusPlus, swin-unet, DoubleUNet, EncoderDoubleCoordAttUNet, and the actual labeled result Mask. Among them, EncoderDoubleCoordAttUNet is the result of this experimental embodiment. The proposed method, a brain vascular endothelial segmentation method based on the dual-coordinate attention mechanism (EncoderDoubleCoordAttUNet), has a VS value of 0.8955, which intuitively verifies the key value of the complementarity of dual-coordinate features in the representation of anatomical structures in medical images—by fusing features from the Cartesian coordinate domain and the polar coordinate domain, it effectively makes up for the shortcomings of single coordinate representation in spatial perception. Sensitivity analysis of the attention deployment stage reveals that when the DCA module (DCA-Decoder) is introduced in all decoder stages, the model performance becomes unstable, with a Dice value of 0.8200 and an HD95 of 10.9822. While these values ​​are higher than the baseline, they are significantly worse than the model with DCA introduced only at the decoder end (DCA-Decoder-Terminal, Dice = 0.8502, HD95 = 7.9957). This result indicates that excessive introduction of attention mechanisms in the decoder stage may disrupt the continuity of cross-scale features. In contrast, introducing DCA only in the final decoder layer achieves stable performance, validating the effectiveness of decoder-end attention in accurately aligning cross-path features and directly optimizing the final segmentation boundary. Furthermore, when DCA (DCA-Encoder) is introduced only in the encoder stage, the model achieves the best overall performance (Dice = 0.8707, HD95 = 6.6880), and also performs best in Sensitivity (0.9053) and IoU (0.7844). This indicates that introducing anatomical priors in the early stages of encoder development is key to improving segmentation performance—encoder-level attention can achieve alignment of dual-path information in the early stages of feature extraction, providing a more stable and reliable structural prior for subsequent decoding. Analysis of the negative impact of excessive attention interaction shows that in the Full DCA model where DCA is introduced in all stages of both encoder and decoder, the overall performance significantly decreases (Dice = 0.7441, HD95 = 12.6649). This phenomenon indicates that excessive interaction of the attention mechanism leads to feature convergence between different paths, weakening the inherent complementary advantages of dual-coordinate representation. Therefore, the deployment of the attention module requires meticulous design and strict control; blindly increasing interaction positions not only fails to improve performance but may also introduce feature redundancy, reducing the model's discriminative ability.

[0074] In summary, the proposed brain vascular endothelial segmentation method based on dual-coordinate attention mechanism (a dual-path dual-coordinate attention method for accurate segmentation of brain vascular endothelial cells in OCT images) aims to address the problems of limited labeled samples, Cartesian-polar coordinate feature mismatch, and insufficient single-coordinate domain attention modeling capability in existing methods. The main contributions include: (1) constructing a high-quality, manually annotated brain vascular OCT dataset to provide a reliable data foundation for model training and evaluation; (2) designing a differentiable polar coordinate transformation module to achieve robust feature alignment between the Cartesian and polar coordinate domains; and (3) proposing a lightweight dual-coordinate attention (DCA) mechanism to support adaptive fusion of complementary structural information in the Cartesian and polar coordinate domains. Extensive experimental results demonstrate that the proposed method significantly outperforms mainstream U-Net-like segmentation models (Dice = 0.8707, HD95 = 6.6880) on key evaluation metrics such as Dice coefficient and HD95, validating the effectiveness of dual-domain feature fusion and cross-coordinate attention mechanisms in brain vascular endothelial segmentation tasks. Future work will extend to spatiotemporal consistency modeling in OCT sequence segmentation, and plans to release some labeled data while adhering to privacy protection principles, in order to promote the research and clinical translation of intelligent analysis methods for cerebrovascular images.

[0075] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for segmenting cerebral vascular endothelium based on a dual-coordinate attention mechanism, characterized in that, Includes the following steps: S1, acquire cerebral vascular optical coherence tomography (OCT) images, perform center alignment, intensity normalization, and data augmentation on the cerebral vascular OCT images to obtain preprocessed images; S2, synchronously input the preprocessed image to the Cartesian coordinate branch and the polar coordinate branch; The Cartesian coordinate branch extracts features from the preprocessed image through a convolution operation and outputs a first feature map. The polar coordinate branch transforms the preprocessed image from the Cartesian coordinate system to the polar coordinate system through a differentiable polar coordinate transformation module, and then extracts features from the polar coordinate image through a convolution operation to output a second feature map. S3, input the first feature map and the second feature map together into the dual-path coordinate attention module, extract one-dimensional directional feature descriptors along the horizontal direction of the Cartesian coordinate system, the vertical direction of the Cartesian coordinate system, the radial direction of the polar coordinate system, and the angular direction of the polar coordinate system, respectively, and perform cross-coordinate domain feature fusion to generate attention weight maps corresponding to the four directions; The dual-path coordinate attention module is deployed in the encoder stage of the U-Net segmentation network and inserted after the skip connections; S4. Based on the attention weight map, perform element-wise multiplication of the first feature map and the second feature map for attention recalibration to obtain the enhanced first output feature map and the second output feature map. Attention recalibration is performed on the first and second feature maps based on the attention weight map, using element-wise multiplication. The calculation formula is as follows: ； ； in, and These are the enhanced first and second output feature maps, respectively. and These are the first feature map and the second feature map, respectively. , , and These are row-direction attention maps, column-direction attention maps, radial attention maps, and angular attention maps, respectively. S5, the enhanced first output feature map and the second output feature map are input into the segmentation network for training and inference, and finally the segmentation result of cerebral vascular endothelium is output; The segmentation results are quantitatively evaluated using region overlap and boundary distance evaluation metrics. The formula for calculating region overlap is as follows: ； ； Where Dice represents the overlap between the prediction and the ground truth annotation; P is the set of predicted segmented pixels, and G is the set of ground truth labeled pixels; Represents the set of boundary points of the predicted segmentation result; Represents the set of boundary points that are actually labeled; express One of the boundary pixels; express One of the boundary pixels; This represents the Euclidean distance between pixels; This means sorting the nearest distances of all boundary points and taking the distance value corresponding to the 95th percentile. This indicates the degree of alignment between the predicted boundary and the true boundary.

2. The brain vascular endothelial segmentation method based on dual-coordinate attention mechanism according to claim 1, characterized in that, In S2, the polar coordinate branch transforms the preprocessed image from Cartesian coordinates to polar coordinates using a differentiable polar coordinate transformation module. The specific transformation method is as follows: The center position and maximum radius of the preprocessed image in the Cartesian coordinate system are determined. In the polar coordinate domain, the center position is used as the pole, and uniform sampling is performed radially within a range not exceeding the maximum radius, and uniform sampling is also performed angularly to construct a polar coordinate sampling grid. Each sampling point in the polar coordinate sampling grid is mapped back to the corresponding coordinate position in the Cartesian coordinate system according to its radial distance and angle. The corresponding coordinate position in the preprocessed image is sampled through bilinear interpolation to obtain the polar coordinate domain feature map.

3. The brain vascular endothelial segmentation method based on dual-coordinate attention mechanism according to claim 1, characterized in that, In S3, one-dimensional directional feature descriptors are extracted, specifically including: Global average pooling is performed on the first feature map along its horizontal and vertical directions to obtain row direction descriptors and column direction descriptors. Global average pooling is performed on the second feature map along its radial and angular directions to obtain radial and angular descriptors, respectively.

4. The brain vascular endothelial segmentation method based on dual-coordinate attention mechanism according to claim 3, characterized in that, In S3, attention weight maps corresponding to the four directions are generated, specifically including: The row direction descriptor, column direction descriptor and radial descriptor are concatenated, processed by the first lightweight bottleneck network to generate the first fusion feature, which is then split into row direction attention map and column direction attention map; The row direction descriptors, column direction descriptors, and angular direction descriptors are concatenated, processed by the second lightweight bottleneck network, and a second fusion feature is generated, which is then split into radial attention maps and angular attention maps.

Images