Retinal vessel topological tree segmentation method and device in wide-angle OCTA and storage medium

Abstract

The invention discloses a retinal vessel topological tree segmentation method and device in wide-angle OCTA and a storage medium. The method comprises the following steps: converting a wide-angle OCTA image into a multi-channel feature map; a coordinate attention module used for enhancing spatial position perception, a cross point attention module used for enhancing blood vessel cross region features and a lightweight ASPP module used for multi-scale feature fusion are embedded in an nnUNetv2 architecture, and a basic segmentation model is constructed; performing end-to-end training on the basic segmentation model by adopting a multi-channel feature map, and in the training process, optimizing network parameters by adopting a composite loss function through back propagation to obtain an OCTA image segmentation model; and based on the OCTA image segmentation model, inputting a wide-angle OCTA image, and outputting a retinal blood vessel segmentation result with a complete topological structure. According to the method, the recognition accuracy of the blood vessel intersection in the wide-angle OCTA image can be improved, and the topological continuity of the blood vessel tree is guaranteed.

Description

Technical Field

[0001] This invention relates to an automatic segmentation and reconstruction method for retinal vessel topology trees in optical coherence tomography (OCTA), and more particularly to a method, device, and storage medium for retinal vessel topology tree segmentation in wide-angle OCTA based on an improved nnUNetv2 architecture. Background Technology

[0002] Accurate segmentation of the retinal vascular topology tree is crucial for the diagnosis of ocular and systemic diseases. Currently, methods based on deep learning models such as U-Net still suffer from the following significant limitations in wide-angle OCTA image segmentation: 1. Poor ability to identify intersections: The proportion of blood vessel intersections is extremely low, and there is a serious class imbalance, which leads to serious missed detections and misjudgments by general models, making it impossible to accurately construct topological connection relationships.

[0003] 2. Weak preservation of topological continuity: Traditional loss functions (such as Dice and cross-entropy) only focus on pixel accuracy and cannot perceive the connectivity of blood vessels, resulting in easy breakage of capillaries in the segmentation results and damage to the integrity of the vascular tree.

[0004] 3. Insufficient feature extraction capability: Existing models have difficulty capturing multi-scale features of thick and thin blood vessels simultaneously and are not sensitive to spatial location information, resulting in unclear segmentation boundaries and inaccurate localization.

[0005] 4. Lack of dedicated structure enhancement: There is a lack of targeted feature enhancement mechanisms for key complex structures such as intersections, which affects the robustness of segmentation. Summary of the Invention

[0006] Objective of this invention: The objective of this invention is to provide a method for retinal vessel topology tree segmentation in wide-angle OCTA images, addressing the issues of improving the accuracy of vessel intersection point recognition and ensuring the topological continuity of the vessel tree. Another objective of this invention is to provide a computer device and storage medium that address the problem of how to execute the retinal vessel topology tree segmentation method in wide-angle OCTA images.

[0007] Technical solution: The present invention provides a method for retinal vessel topology tree segmentation in wide-angle OCTA, comprising the following steps: Convert wide-angle OCTA images into multi-channel feature maps; A basic segmentation model is constructed by embedding a coordinate attention module for enhancing spatial location awareness, a cross-point attention module for strengthening the features of blood vessel crossing regions, and a lightweight ASPP module for multi-scale feature fusion in the nnUNetv2 architecture. The basic segmentation model is trained end-to-end using multi-channel feature maps. A composite loss function is used during training, and the network parameters are optimized through backpropagation to obtain the OCTA image segmentation model. Based on the OCTA image segmentation model, a wide-angle OCTA image is input, and the retinal vessel segmentation result with complete topological structure is output.

[0008] Preferably, the multi-channel feature map includes a grayscale image, a Frangi filter image, and a gradient magnitude image.

[0009] Preferably, the operation of the coordinate attention module includes: The system receives the input feature map and performs position encoding through adaptive pooling in both the horizontal and vertical directions. The encoded features from the two directions are then fused and subjected to convolution and nonlinear transformation to generate spatially aware attention weights. Finally, these weights are multiplied with the original features to output spatially enhanced features.

[0010] Preferably, the operation of the cross-point attention module includes: The system receives the input feature map and performs parallel multi-scale feature extraction through depth-separable dilated convolutions with different dilation rates. The extracted multi-scale features are concatenated along the channel dimension and then subjected to 1×1 convolution dimensionality reduction, batch normalization, and a Sigmoid activation function to generate a spatial attention map. The spatial attention map is then multiplied element-wise with the original input feature map to achieve feature recalibration and output the enhanced features of the blood vessel intersection.

[0011] Preferably, the operation of the lightweight ASPP module includes: The input feature map is processed through parallel processing branches to extract multi-scale features: including a 1×1 convolution branch to preserve the original features, a multi-scale dilated convolution branch to capture the contextual information of different receptive fields, and a global average pooling branch to obtain global features; then the original features, the contextual information of different receptive fields and the global features are concatenated and fused to output the fused and enhanced features of multi-scale vascular features.

[0012] Preferably, the composite loss function is obtained by weighted summation of the Dice loss for vessel intersections, vessels and background, cross-entropy loss and topological loss.

[0013] Preferably, the cross-entropy loss assigns double weight to the cross-point category and ignores the background pixel loss calculation; In the weighted Dice loss, the blood vessel intersections are given a weight of 1.5 times, the blood vessels are given a weight of 1.0 times, and the background weight is 0. In the composite loss function, the weight of topological loss is 2.5, the weight of weighted Dice loss is 1.0, and the weight of cross-entropy loss is 0.3.

[0014] Preferably, the topology loss is based on a differentiable skeleton extraction algorithm, which extracts the centerline through iterative erosion and dilation operations to optimize vessel connectivity and topology.

[0015] A second aspect of the present invention discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0016] A third aspect of the present invention discloses a computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method.

[0017] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention can significantly improve the accuracy of blood vessel intersection identification and reduce missed detections and misjudgments; (2) The present invention can ensure the topological continuity of the vascular tree and avoid segmentation and breakage. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 This is a schematic diagram of the coordinate attention module architecture; Figure 3 The flowchart shows the workflow of the coordinate attention module; Figure 4 This is a schematic diagram of the architecture of the cross-point attention module; Figure 5 A flowchart of the workflow for the cross-point attention module; Figure 6 This is a schematic diagram of the architecture of a lightweight ASPP module. Figure 7 A flowchart of the workflow for the lightweight ASPP module; Figure 8 The three-channel feature map is generated from the original wide-angle OCTA image after preprocessing; Figure 9 This is a comparison image before and after processing the wide-angle OCTA image according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, a method for retinal vessel topology tree segmentation in wide-angle OCTA includes the following steps: (1) Convert the original wide-angle OCTA image into a three-channel feature map (grayscale image, Frangi filter image, gradient magnitude image) to enhance vascular structure information; (2) Embed the CoordAtt module for enhancing spatial location awareness, the CrossNodeAtt module for enhancing the features of the blood vessel crossing region, and the lightweight ASPP module for multi-scale feature fusion into the nnUNetv2 architecture to construct the basic segmentation model. The specific method for embedding the three custom modules into the existing nnUNet model is as follows: First, a MiniASPP module is embedded in the bottleneck layer (after encoder5). Through its five parallel branches (1×1 convolution, three 3×3 dilated convolutions with different dilation rates, and a global average pooling branch), it achieves multi-scale feature fusion, enhancing the network's ability to perceive the different scale structures of blood vessels. Then, a CrossNodeAtt module is embedded in the shallow stage of the decoder (after decoder1). It uses multi-scale dilated convolution (dilation rate [1,2,4,8]) to specifically enhance the attention weights of the blood vessel intersection region, improving the accuracy of intersection detection. Finally, a CoordAtt module is embedded after the output layer. Through the attention mechanism of decomposing the spatial dimension into horizontal and vertical directions, it enhances the spatial position of the final segmentation result, improving the accuracy and continuity of blood vessel boundaries.

[0021] The architecture and working process of the CoordAtt module are as follows: like Figure 2 As shown, the architecture of CoordAtt consists of the following steps: First, the original features are preserved through the residual branch. At the same time, global average pooling is performed on the input in the X direction (along the width) and Y direction (along the height) in the side path to obtain two positional features, C×H×1 and C×1×W. These features are then concatenated in the spatial dimension to form C×1×(W+H). After compression of the channels to C / r by 1×1 convolution, batch normalization and non-linear activation are performed, and then the features are decomposed back to C×H×1 and C×1×W. Each feature is then decomposed into C×H×1 and C×1×W by 1×1 convolution and then processed by Sigmoid to generate horizontal and vertical attention weights. Finally, these two one-dimensional weights are broadcast and multiplied back into the residual features to achieve a lightweight spatial-channel joint enhancement that preserves coordinate information. The output is an enhanced feature map of the same size as the input.

[0022] The working process of CoordAtt is as follows: Figure 3 As shown, the system receives the input feature map and performs position encoding through adaptive pooling in both the horizontal and vertical directions. The encoded features from both directions are then fused and subjected to convolution and nonlinear transformations to generate spatially aware attention weights. Finally, these weights are multiplied by the original features to output spatially enhanced features. This mechanism effectively improves the model's ability to perceive the spatial distribution and topological relationships of blood vessels. This module can capture spatial location information and improve the accuracy of blood vessel localization. like Figure 4 As shown, the architecture of the CrossNodeAtt module is as follows: After the input features are reduced in dimensionality by a common 1×1 convolution, they are passed in parallel through four 3×3 depthwise separable dilated convolutions (dilation=1,2,4,8) to form a cross-shaped multi-scale receptive field in the horizontal and vertical directions. The outputs of the four branches are concatenated into a 4C×H×W structure in the channel dimension, and then reduced back to the C channel by a 1×1 convolution and batch normalized + Sigmoid to generate an attention weight map of [0,1]. Finally, the attention weights are multiplied element-wise with the original input by residual skip connections to achieve a lightweight spatial-channel joint enhancement that preserves coordinate information. The output is an enhanced feature map of the same size as the input and rich in cross-point features.

[0023] CrossNodeAtt workflow as follows Figure 5 As shown, this module is specifically designed for vascular intersections. After receiving the input feature map, it performs parallel multi-scale feature extraction through depthwise separable dilated convolutions with different dilation rates. The extracted multi-scale features are concatenated along the channel dimension, then subjected to 1×1 convolution dimensionality reduction, batch normalization, and a sigmoid activation function to generate a spatial attention map. The spatial attention map is then multiplied element-wise with the original input feature map to achieve feature recalibration, outputting enhanced features for vascular intersections. The specific workflow is as follows: Input feature map (B=batch size, C=channels, H=height, W=width) The input is processed by a set of dilated convolutional layers operating in parallel. Each convolutional layer uses a different dilation rate d∈{1,2,4,8} and adopts a depthwise separable convolutional form, that is, the number of groups is set to C, so that each input channel is processed independently.

[0024] For each inflation rate d, the convolution operation is represented as:

[0025] Output feature map Next, all the dilated convolution outputs are concatenated along the channel dimension to obtain:

[0026] Subsequently, multi-scale features are fused through a 1×1 convolutional layer, the weights of which are... The calculation is as follows:

[0027] Then, batch normalization is applied, and an attention map is generated using the Sigmoid activation function. Finally, the module passes through... The feature map of the original input is multiplied element-wise with the attention map to obtain the enhanced feature of the blood vessel intersection.

[0028] like Figure 6 As shown, the architecture of the lightweight ASPP module (MiniASPP) is as follows: First, dimensionality reduction is performed using 1×1 convolution, followed by four parallel branches: ordinary 3×3 convolution (d=1), two sets of dilated convolutions (d=2,3) and GAP-upsampling global branch. The four outputs are concatenated in the channel dimension to form 4C×H×W. Then, 1×1 convolution is used to fuse the 4C back to C, and BN+ReLU is used to obtain an enhanced feature map of the same size as the input, rich in multi-scale context.

[0029] like Figure 7 As shown, the working process of MiniASPP is as follows: The input feature map is processed through parallel processing branches to extract multi-scale features: a 1×1 convolution branch to preserve the original features, a multi-scale dilated convolution branch to capture contextual information from different receptive fields, and a global average pooling branch to obtain global features. The original features, contextual information from different receptive fields, and global features are then concatenated and fused to output a fused and enhanced multi-scale vascular feature. This module can fuse multi-scale contextual information to enhance the model's perceptual capabilities.

[0030] (3) A modular training process for enhancing the basic segmentation model using multi-channel feature maps is as follows: First, the original training data is processed in three channels. Then, the basic segmentation model after embedding the module is used for end-to-end training. Finally, the prediction file is visualized to obtain the final result.

[0031] A composite loss function is used during training, and the network parameters are optimized through backpropagation to obtain the OCTA image segmentation model; The composite loss function is designed as follows: Cross-entropy loss (CE): Double the weight of the intersection category and ignore the loss calculation of background pixels; Three-class weighted Dice loss (DiceLoss): 1.5 times weight is given to the blood vessel intersection, 1.0 times weight is given to the blood vessel, and the background weight is 0; Topological loss (SoftClDiceLoss): Based on a differentiable skeleton extraction algorithm, it extracts the centerline through iterative erosion and dilation operations, specifically optimizing vessel connectivity and topological structure. The specific process is as follows: First, the predicted probabilities of blood vessels (category 1) and intersections (category 2) are summed and clipped to the range [0,1] to obtain the soft prediction mask:

[0032] Where P :,1 P represents the probability value of class 1 (blood vessels) across all samples and all spatial locations. :,2This represents the probability value for category 2 (intersection). A binary vascular mask is also generated from the ground truth labels.

[0033] Where 1{·} is an indicator function (returns 1 if the condition is true, otherwise returns 0), y represents the true label (0 for background, 1 for blood vessel, 2 for blood vessel intersection), and the condition y=1∨y=2 checks whether the label is a blood vessel or an intersection. Then, differentiable soft skeletonization operations are performed on both: by iteratively applying soft erosion and soft expansion 10 times each, the formula is:

[0034] in, max_pool2d(-x) This indicates that max pooling is performed on the negative feature map.

[0035] Soft expansion formula:

[0036] Extracting centerline features yields P skel (Predicted skeleton map) and T skel (Actual skeleton diagram)

[0037] Then, the topological accuracy is calculated using the following formula:

[0038] This formula indicates the proportion of the predicted skeleton that is actually located within the vascular region. The closer the value is to 1, the more accurate the predicted skeleton is and the fewer false branches there are.

[0039] The formula for calculating topological sensitivity is:

[0040] This formula indicates the proportion of the actual skeletal structure covered by the predicted vascular region. The closer the value is to 1, the more complete the actual vascular structure is detected, and the fewer the missed detections.

[0041] The final clDice value is the harmonic mean of the two values, calculated using the following formula:

[0042] When both topology accuracy and topology sensitivity are high, the clDice value is close to 1, indicating that the predicted vascular skeleton is highly consistent with the real vascular skeleton in terms of topology. When one of these metrics is low, the clDice value drops significantly, emphasizing that both must be optimized simultaneously. This ensures the topological integrity of the vascular network, reducing false branches (false positives) and avoiding the omission of true branches (false negatives).

[0043] The loss function is defined as L clDice =1-Dice When clDice is close to 1, the loss is close to 0, indicating that the model predicts a good topology. During training, the model minimizes the loss through gradient descent. L clDice This indirectly maximizes clDice and improves the topological quality of vessel segmentation.

[0044] The formula for the composite loss function is: TotaLoss=0.3*CE+1.0*DiceLoss+2.5*SoftClDiceLoss The weight allocation is as follows: SoftClDiceLoss weight 2.5 (topology priority), DiceLoss weight 1.0 (basic accuracy), and cross-entropy loss weight 0.3 (class balance).

[0045] (4) Based on the OCTA image segmentation model, input a wide-angle OCTA image and output a retinal vessel segmentation result with complete topological structure.

[0046] Implementation example: Taking the WF-OCTA80 dataset as an example, the specific implementation steps are as follows: Data preparation: Use 60 training images and 20 test images, each image is 20mm×24mm in size.

[0047] Preprocessing: Generate a three-channel feature map: Channel 0 is a grayscale image, Channel 1 is a Frangi-filtered image, and Channel 2 is a gradient magnitude image (e.g., ...). Figure 8 (as shown) Label conversion: background is 0, ordinary blood vessels are 1, and intersections are 2.

[0048] Training the basic segmentation model: An improved nnUNetv2 architecture with embedded CoordAtt, CrossNodeAtt, and lightweight ASPP modules is used for end-to-end training with a composite loss function. Reasoning and Application: Input a wide-angle OCTA image (such as...) Figure 9 (As shown in the middle left image), output the segmentation result image, as follows: Figure 9 As shown in the middle right image, the intersections are set to yellow, the blood vessels to red, and the background to black.

[0049] The segmentation results output by this invention can be used for early screening of diseases such as diabetic retinopathy and glaucoma, or as a quantitative indicator for risk assessment of systemic diseases.

[0050] This invention can also be applied to other medical image segmentation tasks, such as topological reconstruction of brain blood vessels, lung blood vessels, etc., only requiring adjustments to data preprocessing and label definitions.

Claims

1. A method for retinal vessel topology tree segmentation in wide-angle OCTA, characterized in that, Includes the following steps: Convert wide-angle OCTA images into multi-channel feature maps; A basic segmentation model is constructed by embedding a coordinate attention module for enhancing spatial location awareness, a cross-point attention module for strengthening the features of blood vessel crossing regions, and a lightweight ASPP module for multi-scale feature fusion in the nnUNetv2 architecture. The basic segmentation model is trained end-to-end using multi-channel feature maps. A composite loss function is used during training, and the network parameters are optimized through backpropagation to obtain the OCTA image segmentation model. Based on the OCTA image segmentation model, a wide-angle OCTA image is input, and the retinal vessel segmentation result with complete topological structure is output.

2. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 1, characterized in that, The multi-channel feature map includes a grayscale image, a Frangi filter image, and a gradient magnitude image.

3. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 1, characterized in that, The operation of the coordinate attention module includes: The system receives the input feature map and performs position encoding through adaptive pooling in both the horizontal and vertical directions. The encoded features from the two directions are then fused and subjected to convolution and nonlinear transformation to generate spatially aware attention weights. Finally, these weights are multiplied with the original features to output spatially enhanced features.

4. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 1, characterized in that, The working process of the cross-point attention module includes: The system receives the input feature map and performs parallel multi-scale feature extraction through depth-separable dilated convolutions with different dilation rates. The extracted multi-scale features are concatenated along the channel dimension and then subjected to 1×1 convolution dimensionality reduction, batch normalization, and a Sigmoid activation function to generate a spatial attention map. The spatial attention map is then multiplied element-wise with the original input feature map to achieve feature recalibration and output the enhanced features of the blood vessel intersection.

5. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 1, characterized in that, The operation of the lightweight ASPP module includes: The input feature map is processed through parallel processing branches to extract multi-scale features: including a 1×1 convolution branch to preserve the original features, a multi-scale dilated convolution branch to capture the contextual information of different receptive fields, and a global average pooling branch to obtain global features; then the original features, the contextual information of different receptive fields and the global features are concatenated and fused to output the fused and enhanced features of multi-scale vascular features.

6. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 1, characterized in that, The composite loss function is obtained by weighted summation of the Dice loss for vessel intersections, vessels, and background, the cross-entropy loss, and the topological loss.

7. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 6, characterized in that, In the cross-entropy loss, the category of the cross point is given double weight, and the loss calculation of background pixels is ignored. In the weighted Dice loss, the blood vessel intersections are given a weight of 1.5 times, the blood vessels are given a weight of 1.0 times, and the background weight is 0. In the composite loss function, the weight of topological loss is 2.5, the weight of weighted Dice loss is 1.0, and the weight of cross-entropy loss is 0.

3.

8. The retinal vessel topology tree segmentation method in wide-angle OCTA according to claim 6, characterized in that, The topology loss is based on a differentiable skeleton extraction algorithm, which extracts the centerline through iterative erosion and dilation operations to optimize blood vessel connectivity and topology.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed on the computer, cause the computer to perform the method as described in any one of claims 1-8.

Images