OCTA image retinal vessel segmentation method based on full-resolution network

Abstract

The invention provides an OCTA image retinal vessel segmentation method based on a full-resolution network, and the method comprises the steps: carrying out the preprocessing of an input OCTA image through a data preprocessing module, wherein the preprocessing comprises the optical transformation and geometric transformation, and obtaining the preprocessed data; the feature coding module uses five feature extraction units which are arranged in sequence to extract image features of five levels and outputs the image features to the cascade feature enhancement module; a cascade feature enhancement module CFEM combines low-level detail information and high-level semantic information through cross-layer guidance for feature fusion, and then outputs a segmentation result; calculating loss and optimizing the OCTA image retinal vessel segmentation model based on the full-resolution network to obtain an optimized segmentation model; using the optimized segmentation model to obtain a retinal blood vessel image; according to the method, high-precision and high-efficiency segmentation of the retinal blood vessel can be realized, the network complexity is relatively low, and the computing resource consumption is relatively low.

Description

Technical Field

[0001] The present invention relates to a full-resolution network-based OCTA image retinal blood vessel segmentation method, belonging to the technical field of image segmentation. Background Art

[0002] With the rapid development of science and technology, automatic segmentation of retinal blood vessels has become increasingly important in medical image analysis. Existing methods for blood vessel segmentation have limitations. For example, they are not ideal for segmenting low-contrast and small blood vessels. Traditional methods such as thresholding, edge detection, and morphological operations perform poorly in complex backgrounds and low-contrast images.

[0003] In recent years, some progress has been made in methods for retinal vessel segmentation in OCTA images. For example, a global fusion dual-channel retinal vessel segmentation method preprocesses and resamples retinal vessel images, feeds images of different sizes into parallel small vessel enhancement and segmentation networks, and ultimately trains them using a specific loss function. While this method performs well in segmenting small vessels, its network structure is complex, training time is long, and it has certain limitations when dealing with complex backgrounds.

[0004] Another approach is a deep learning-based retinal vessel segmentation method for OCTA images, proposed by Shanghai Normal University. This method utilizes a dual-branch axially compressed convolutional neural network segmentation model. This model enhances segmentation performance through multi-scale image input and cross-feature fusion. While this method has improved accuracy and processing speed, it still faces challenges when processing complex backgrounds and low-contrast images. Furthermore, the network structure is complex, resulting in high computational resource consumption.

[0005] There are also coronary artery segmentation methods based on transfer learning. This method utilizes transfer learning to improve segmentation accuracy by pre-training a neural network and a dense semantic supervision module, thereby addressing the scarcity of coronary artery data. However, the application of transfer learning in retinal vessel segmentation requires further verification, and its network design is complex and requires significant computational resources.

[0006] Although the above methods have improved the performance of retinal vessel segmentation to varying degrees, there are still some problems such as high network complexity, long processing time, and poor adaptability to low-contrast images. Summary of the Invention

[0007] The purpose of the present invention is to provide a full-resolution network-based OCTA image retinal vessel segmentation method to solve the problems of high network complexity, large computing resource consumption and low segmentation accuracy in the prior art.

[0008] The technical solution of the present invention is: A full-resolution network-based OCTA image retinal vessel segmentation method includes the following steps: S1. The input OCTA image is preprocessed by the data preprocessing module, including optical transformation and geometric transformation, to obtain preprocessed data; S2. The pre-processed data is input into a feature encoding module, which uses five sequentially arranged feature extraction units to extract five levels of image features and outputs them to a cascade feature enhancement module; S3, the cascade feature enhancement module CFEM combines low-level detail information with high-level semantic information through cross-layer guidance to perform feature fusion and output the segmentation result; S4, constructing a full-resolution network-based OCTA image retinal vessel segmentation model using a data preprocessing module, a feature encoding module, and a cascade feature enhancement module, calculating the loss and optimizing the full-resolution network-based OCTA image retinal vessel segmentation model to obtain an optimized segmentation model; S5. Use the optimized segmentation model to obtain a retinal vascular image on the OCTA image to be segmented.

[0009] Furthermore, in step S1, the optical transformation includes randomly adjusting contrast, hue and saturation, and the geometric transformation includes random expansion, random cropping, random mirroring and scaling to a fixed ratio.

[0010] Furthermore, in step S2, each feature extraction unit includes an improved recurrent convolutional network block, namely an improved RLConvNeXt block and a similarity-based attention mechanism layer SimAM. Improved RLConvNeXt block: extracts multi-dimensional channel features through convolutions and residual connections of different dimensions; Similarity-based attention mechanism layer SimAM: The multi-dimensional channel features of the input are used to capture features and contextual information at different levels and then output.

[0011] Furthermore, the improved RLConvNeXt block includes an enhanced local attention layer ELA, a deep convolutional layer DwConv, a LayerNorm layer, a convolutional layer Conv1, a GELU layer and a convolutional layer Conv2. After the input features pass through the ELA layer, they are input into the deep convolutional layer DwConv to capture the spatial features of the image. After that, they pass through the LayerNorm layer and the convolutional layer Conv1 in sequence and then undergo a nonlinear transformation by the GELU layer. After entering the convolutional layer Conv2, the convolved features are output and residually connected with the input features to obtain the output features.

[0012] Furthermore, in step S3, the cascade feature enhancement module includes a cross-layer guidance module, a hybrid attention module and a multi-axis attention module. Cross-layer guidance module: Through skip connections, the image features of the 1st and 2nd layers (i.e., low-level detail features) output by the feature encoding module are fused with the image features of the 3rd to 5th layers (i.e., high-level semantic features) and then output to the hybrid attention module; Hybrid Attention Module: It includes five hybrid attention layers. Each hybrid attention layer highlights important features through spatial attention and channel attention mechanisms and outputs them to the multi-axis attention module. Multi-axis attention module: Five multi-axis attention blocks (GMA Blocks) are used, and each mixed attention layer is input into each multi-axis attention block (GMA Block) one by one. Each multi-axis attention block (GMA Block) captures multi-scale features from multiple spatial dimensions, and the output features of the five multi-axis attention blocks (GMA Blocks) are vector-added and output.

[0013] Furthermore, in the hybrid attention module, the first hybrid attention layer Hybrid Attn1 highlights the important features of the first-level image features of the input and outputs them; the second hybrid attention layer Hybrid Attn2 highlights the important features of the input fusion feature one and outputs them; the third hybrid attention layer Hybrid Attn3 highlights the important features of the input fusion feature two and outputs them; the fourth hybrid attention layer Hybrid Attn4 highlights the important features of the input fusion feature three and outputs them; the fifth hybrid attention layer Hybrid Attn5 highlights the important features of the input fusion feature four and outputs them.

[0014] Furthermore, the hybrid attention layer includes a channel attention mechanism module, a spatial attention mechanism module and a residual connection layer. In the channel attention mechanism module, the input image features pass through the average pooling layer 1 Avg Pool1 and are output to Reshape1 and average pooling layer 2 Avg Pool2 respectively. The output features of the reshaping layer 1 Reshape1 pass through the convolution layer 1 Conv1 and are output to the reshaping layer 2 Reshape2. The output features of the average pooling layer 2 Avg Pool2 pass through the reshaping layer 3 Reshape3, the convolution layer 2 Conv2, and the average pooling layer 3 Avg Pool3 in sequence. The output features of the reshaping layer 2 Reshape2 are residually connected with the output features of the average pooling layer 3 Avg Pool3 and then output to the average pooling layer 4 Avg Pool4. The output features of the average pooling layer 4 Avg Pool4 are weighted with the input image features channel by channel and then output. In the spatial attention mechanism module, the input image features are respectively average pooled along the X-axis by the average pooling layer X Avg Pool and average pooled along the Y-axis by the average pooling layer Y Avg Pool. The output features of the average pooling layer X Avg Pool are sequentially passed through the convolution layer three Conv3, the group normalization layer GroupNorm1, and the activation function layer Sigmoid1 to output the X-direction spatial weight map. The output features of the average pooling layer Y Avg Pool are sequentially passed through the convolution layer four Conv4, the group normalization layer GroupNorm2, and the activation function layer Sigmoid2 to output the Y-direction spatial weight map. The X-direction spatial weight map and the Y-direction spatial weight map are multiplied point by point and then multiplied with the input image features for output. Residual connection layer: The output features of the channel attention mechanism module are residually connected with the output features of the spatial attention mechanism module and then output.

[0015] Furthermore, the multi-axis attention block GMA Block includes LayerNorm layer 1, 2D separable convolution layer, 1D convolution layer 1, 1D convolution layer 2, depthwise separable convolution layer, splicing layer and LayerNorm layer 2. The input image feature map is normalized by LayerNorm layer 1 and then output to the two-dimensional separable convolution layer, one-dimensional convolution layer 1, one-dimensional convolution layer 2, and depth-wise separable convolution layer; 2D separable convolution layer: As the XY branch, it uses 2D separable convolution to process the original spatial structure of the input feature map, models the spatial attention on the XY axis plane, and outputs it to the splicing layer; One-dimensional convolution layer 1: As the ZX branch, it transposes the input feature map in the width dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; One-dimensional convolution layer 2: As the ZY branch, it transposes the input feature map in the height dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; Depthwise separable convolutional layer: extracts local spatial context information from the input feature map and outputs it to the splicing layer; Splicing layer: Splicing is performed on the channel dimension and then output to the LayerNorm layer 2; Furthermore, in step S5, the loss is calculated using the DiceJaccardLoss loss function which is a weighted combination of the Dice loss and the Jaccard loss.

[0016] The beneficial effects of the present invention are: 1. This full-resolution network-based OCTA image retinal vessel segmentation method, by utilizing an improved full-resolution network, can achieve high-precision and efficient segmentation of retinal vessels, especially in processing low-contrast and small vessels. It also has low network complexity and consumes less computing resources.

[0017] 2. This full-resolution network-based OCTA image retinal vessel segmentation method can achieve efficient and accurate vessel segmentation through the improved RLConvNeXt block and cascade feature enhancement module CFEM, improve the model's segmentation performance in small blood vessels and low-contrast areas, and significantly improve the accuracy and efficiency of segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1 is a flow chart of a method for retinal vessel segmentation in OCTA images based on a full-resolution network according to an embodiment of the present invention; Figure 2 is a schematic diagram illustrating a full-resolution network-based OCTA image retinal vessel segmentation model in an embodiment; Figure 3 is a schematic diagram illustrating the improved RLConvNeXt block in the embodiment; Figure 4 is a schematic diagram illustrating a hybrid attention layer in an embodiment; Figure 5 Schematic diagram illustrating the multi-axis attention block GMA Block in the embodiment. DETAILED DESCRIPTION

[0019] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0020] The embodiment provides a full-resolution network-based OCTA image retinal vessel segmentation method, such as Figure 1 , including the following steps, S1. The input OCTA image data preprocessing module Stem performs preprocessing including optical transformation and geometric transformation to obtain preprocessed data.

[0021] In step S1, data enhancement is achieved by performing optical transformation and geometric transformation. Optical transformation involves randomly adjusting contrast, hue, and saturation to increase the generalization ability of the model; geometric transformation involves random expansion, random cropping, random mirroring, and scaling to a fixed ratio. Several fixed scales are predefined, and one scale is randomly selected for training in each training cycle to enhance the robustness of the model. At the same time, an edge filling strategy is used during scaling, that is, the edge of the image is filled with a background image. The random expansion operation in the geometric transformation can increase the size of the image by adding random pixels to the edge of the image, random cropping is performed by randomly selecting a sub-region from the image for training, and random mirroring is performed by flipping the image horizontally or vertically to increase the diversity of the training data.

[0022] S2. The pre-processed data is input into a feature encoding module. The feature encoding module uses five feature extraction units arranged in sequence to extract five levels of image features and outputs them to a cascade feature enhancement module.

[0023] In step S2, each feature extraction unit includes an improved recurrent convolutional network block, namely an improved RLConvNeXt block and a similarity-based attention mechanism layer SimAM. Improved RLConvNeXt block: extracts multi-dimensional channel features through convolutions and residual connections of different dimensions.

[0024] Similarity-based attention mechanism layer SimAM: The multi-dimensional channel features of the input are used to capture features and contextual information at different levels and then output.

[0025] like Figure 3 The improved RLConvNeXt block includes an enhanced local attention layer ELA, a deep convolutional layer DwConv, a LayerNorm layer, a convolutional layer Conv1, a GELU layer, and a convolutional layer Conv2. After the input features pass through the ELA layer, they are input into the deep convolutional layer DwConv to capture the spatial features of the image. After that, they pass through the LayerNorm layer and the convolutional layer Conv1 in sequence and then undergo a nonlinear transformation by the GELU layer. After entering the convolutional layer Conv2, the convolutional features are output and residually connected with the input features to obtain the output features.

[0026] In the improved RLConvNeXt block, the deep convolution layer DwConv uses a 7×7 convolution kernel, and the convolution layer 1 Conv1 and the convolution layer 2 Conv2 use a 3×3 convolution kernel. The deep convolution layer DwConv is used to capture the spatial features of the image. This larger convolution kernel can effectively extract local and global information. The convolution layer 1 Conv1 and the convolution layer 2 Conv2 expand the feature space through 1×1 point convolution, thereby increasing the dimension of the feature and enabling the network to learn more complex features. The GELU activation function is used for nonlinear transformation to better maintain gradient flow and reduce the problem of gradient disappearance. The input is directly added to the output through the residual connection, thereby maintaining the original information in the deep network, promoting gradient flow and effective transmission, enhancing feature extraction capabilities, and ensuring the stability of training and improving model performance. The backbone network structure is as follows Figure 3 As shown in the figure, by using direct paths and convolutional layers to perform element-wise addition instead of channel splicing, information flow between feature map channels is ensured. During training, only the residual values ​​between the output and input values ​​in different layers of the network need to be considered, which reduces the difficulty of model learning and helps the model converge quickly.

[0027] S3. The cascade feature enhancement module CFEM combines low-level detail information with high-level semantic information through cross-layer guidance to perform feature fusion and output the segmentation result.

[0028] In step S3, the cascade feature enhancement module includes a cross-layer guidance module, a hybrid attention module and a multi-axis attention module. Cross-layer guidance module: Through skip connections, the image features of the 1st and 2nd levels output by the feature encoding module, namely low-level detail features such as blood vessel edges, are integrated with the image features of the 3rd to 5th levels, namely high-level semantic features such as the overall morphology of blood vessels. Specifically, feature interaction is achieved through element addition or channel splicing.

[0029] Hybrid Attention Module: This module includes five hybrid attention layers. Each hybrid attention layer highlights important features through spatial attention (focusing on blood vessel location) and channel attention mechanism (focusing on feature importance) and outputs them to the multi-axis attention module. In the hybrid attention module, the first hybrid attention layer Hybrid Attn1 highlights the important features of the first-level image features of the input and outputs them; the second hybrid attention layer Hybrid Attn2 highlights the important features of the input fusion feature one and outputs them; the third hybrid attention layer Hybrid Attn3: highlights the important features of the input fusion feature two and outputs them; the fourth hybrid attention layer Hybrid Attn4 highlights the important features of the input fusion feature three and outputs them; the fifth hybrid attention layer Hybrid Attn5 highlights the important features of the input fusion feature four and outputs them.

[0030] like Figure 4 ,The hybrid attention layer includes a channel attention mechanism module, a spatial attention mechanism module and a residual connection layer.

[0031] In the channel attention mechanism module, the input image features pass through the average pooling layer 1 Avg Pool1 and are output to Reshape1 and the average pooling layer 2 Avg Pool2 respectively. The output features of the reshaping layer 1 Reshape1 pass through the convolution layer 1 Conv1 and are output to the reshaping layer 2 Reshape2. The output features of the average pooling layer 2 Avg Pool2 pass through the reshaping layer 3 Reshape3, the convolution layer 2 Conv2, and the average pooling layer 3 Avg Pool3 in turn. The output features of the reshaping layer 2 Reshape2 are residually connected with the output features of the average pooling layer 3 Avg Pool3 and then output to the average pooling layer 4 Avg Pool4. The output features of the average pooling layer 4 Avg Pool4 are weighted channel by channel with the input image features and then output. Among them, the convolution layer 1 Conv1 and the convolution layer 2 Conv2 both use a 3×3 convolution kernel.

[0032] In the spatial attention mechanism module, the input image features are average pooled along the X-axis by the average pooling layer X Avg Pool and average pooled along the Y-axis by the average pooling layer Y Avg Pool. The output features of the average pooling layer X Avg Pool are sequentially passed through the convolution layer three Conv3, the group normalization layer GroupNorm1, and the activation function layer Sigmoid1 to output the X-direction spatial weight map. The output features of the average pooling layer Y Avg Pool are sequentially passed through the convolution layer four Conv4, the group normalization layer GroupNorm2, and the activation function layer Sigmoid2 to output the Y-direction spatial weight map. The X-direction spatial weight map and the Y-direction spatial weight map are point-by-point multiplied and then multiplied with the input image features for output. Among them, the convolution layer three Conv3 and the convolution layer four Conv4 both use a 7×7 convolution kernel.

[0033] Residual connection layer: The output features of the channel attention mechanism module are residually connected with the output features of the spatial attention mechanism module and then output.

[0034] Figure 4In the figure, the branch structure on the left constitutes the channel attention mechanism module, while the branch structure on the right constitutes the spatial attention mechanism module. The channel attention mechanism uses multiple average pooling and Conv 3×3 convolutions, combined with reshape operations and residual connections, to highlight the importance of key channels in the input features and weight them channel by channel. The spatial attention mechanism first performs average pooling on the input feature map along the X and Y axes. After Conv 7×7 convolution, GroupNorm, and Sigmoid activation, it generates spatial weight maps in two directions. These maps are point-by-point multiplied and then multiplied with the input features to achieve saliency enhancement in the spatial dimension. Both are multiplied with the input features and the residuals are added together, improving the representation of vascular structures in OCTA images and enhancing segmentation results.

[0035] Multi-axis attention module: Five multi-axis attention blocks (GMA Blocks) are used, and each mixed attention layer is input into each multi-axis attention block (GMA Block) one by one. Each multi-axis attention block (GMA Block) captures multi-scale features from multiple spatial dimensions, and the output features of the five multi-axis attention blocks (GMA Blocks) are vector-added and output.

[0036] like Figure 5 The multi-axis attention block GMA Block includes LayerNorm layer 1, two-dimensional separable convolution layer, one-dimensional convolution layer 1, one-dimensional convolution layer 2, depth-wise separable convolution layer, splicing layer and LayerNorm layer 2. The input image feature map is normalized by LayerNorm layer 1 and then output to the two-dimensional separable convolution layer, one-dimensional convolution layer 1, one-dimensional convolution layer 2, and depth-wise separable convolution layer; 2D separable convolution layer: As the XY branch, it uses 2D separable convolution to process the original spatial structure of the input feature map, models the spatial attention on the XY axis plane, and outputs it to the splicing layer; One-dimensional convolution layer 1: As the ZX branch, it transposes the input feature map in the width dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; One-dimensional convolution layer 2: As the ZY branch, it transposes the input feature map in the height dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; Depthwise separable convolutional layer: extracts local spatial context information from the input feature map and outputs it to the splicing layer; Splicing layer: Splicing is performed on the channel dimension and then output to the LayerNorm layer 2; LayerNorm layer 2: After normalization, it outputs a feature map that integrates multi-dimensional spatial attention information.

[0037] The Multi-Axis Attention (GMA) Block is an attention mechanism that incorporates multidimensional spatial dependencies, designed to enhance the expressiveness of image features. The GMA Block divides the input feature channels into four groups. Three of these groups model spatial attention on the XY, ZX, and ZY planes, achieving directional saliency enhancement through learnable parameters and element-wise multiplication. The other group extracts local spatial features through depthwise separable convolution. Finally, the four branches are concatenated and their outputs are integrated, effectively capturing spatial dependencies across different dimensions and improving the segmentation accuracy of tiny structures in medical images.

[0038] In step S3, the cascaded feature enhancement module (CFEM) combines low-level detail information with high-level semantic information through cross-layer guidance to improve feature representation. Specifically, low-level features contain rich detailed information, while high-level features provide global contextual semantics. Through cross-layer guidance, these features can be effectively fused at different levels. The hybrid attention module further enhances the selective representation of features. By using attention mechanisms across different spatial and channel dimensions, it effectively highlights important features and suppresses irrelevant information. The multi-axis attention block (GMA Block) captures multi-scale and multi-directional characteristics of blood vessels (such as small vessels with different orientations) through feature computation along the X and Y axes (e.g., axially compressed convolution), effectively improving feature fusion. By performing feature computation along multiple axes, the multi-axis attention block captures multi-scale and multi-directional features in the image, further improving the accuracy and robustness of feature representation. Feature fusion is performed using the hybrid attention module and the multi-axis attention block to enhance feature representation. The hybrid attention module generates attention weights through average pooling and convolution operations, suppressing background noise and enhancing vascular feature responses. The features are further fused and enhanced through the cascade feature enhancement module CFEM, and finally high-quality segmentation results are output. The features of different levels and attention modules are integrated together to form a comprehensive feature representation, such as Figure 4 These comprehensive representations contain not only rich details but also global semantic information, effectively improving the feature expression capability and ensuring that the network can accurately segment retinal vessels in complex visual scenes, especially in small vessels and low-contrast areas.

[0039] In step S3, features are further fused and enhanced through a cascaded feature enhancement module, ultimately outputting a high-quality segmentation result. During the feature fusion process, features from different layers and attention modules are combined to form a comprehensive feature representation that incorporates not only rich detail information but also global semantic information. This fusion approach ensures that the network can accurately segment retinal vessels in complex visual scenes, particularly in small vessels and low-contrast areas.

[0040] S4. A retinal vascular segmentation model for OCTA images based on a full-resolution network is constructed by a data preprocessing module, a feature encoding module, and a cascade feature enhancement module. The loss is calculated and the retinal vascular segmentation model for OCTA images based on a full-resolution network is optimized to obtain an optimized segmentation model.

[0041] S5. Use the optimized segmentation model to obtain a retinal vascular image on the OCTA image to be segmented.

[0042] In step S5, the loss is calculated using the DiceJaccardLoss loss function, which is a weighted combination of Dice loss and Jaccard loss. Dice loss focuses on the overlapping area between the prediction and the true label, while Jaccard loss further considers the ratio of the overlapping area to the joint area. Combining these two loss functions can comprehensively measure the quality of the segmentation results, enhance the segmentation performance of small blood vessels and low-contrast areas, and significantly improve the accuracy and efficiency of retinal vessel segmentation. In addition, the loss function optimization also introduces difficult sample mining. By dynamically adjusting the loss weight, the model can more effectively learn the characteristics of difficult samples, thereby further improving segmentation accuracy and robustness.

[0043] This full-resolution network-based OCTA image retinal vessel segmentation method, by utilizing an improved full-resolution network, can achieve high-precision and efficient segmentation of retinal vessels, especially in processing low-contrast and small blood vessels. It also has low network complexity and consumes less computational resources.

[0044] This full-resolution network-based retinal vessel segmentation method for OCTA images addresses the difficulties and poor segmentation effects of existing image segmentation algorithms. First, an improved RLConvNeXt backbone network is designed and adopted for feature extraction. Effective feature extraction is achieved by adding residual connections and an attention mechanism. Then, semantic information at different levels is fused through a cascaded feature enhancement module (CFEM), effectively combining feature information at different levels. Important information is extracted and retained using the attention mechanism. Finally, the DiceJaccardLoss loss function is used for loss calculation to improve the accuracy and robustness of retinal vessel segmentation.

[0045] This full-resolution network-based retinal vessel segmentation method for OCTA images, through an improved RLConvNeXt block and a cascaded feature enhancement module (CFEM), achieves efficient and accurate vessel segmentation. It improves the model's segmentation performance for small vessels and low-contrast areas, significantly enhancing segmentation accuracy and efficiency. Multi-scale feature fusion and an attention mechanism effectively improve segmentation robustness and accuracy, making it suitable for a variety of medical image analysis scenarios.

[0046] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A full-resolution network-based retinal vessel segmentation method for OCTA images, characterized by: The following steps are included: S1. The input OCTA image is preprocessed by the data preprocessing module, including optical transformation and geometric transformation, to obtain preprocessed data; S2. The pre-processed data is input into a feature encoding module, which uses five sequentially arranged feature extraction units to extract five levels of image features and outputs them to a cascade feature enhancement module; S3, the cascade feature enhancement module CFEM combines low-level detail information with high-level semantic information through cross-layer guidance to perform feature fusion and output the segmentation result; S4, constructing a full-resolution network-based OCTA image retinal vessel segmentation model using a data preprocessing module, a feature encoding module, and a cascade feature enhancement module, calculating the loss and optimizing the full-resolution network-based OCTA image retinal vessel segmentation model to obtain an optimized segmentation model; S5. Use the optimized segmentation model to obtain a retinal vascular image on the OCTA image to be segmented.

2. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 1, characterized in that: In step S1 , the optical transformation includes random adjustment of contrast, hue, and saturation, and the geometric transformation includes random expansion, random cropping, random mirroring, and scaling to a fixed ratio.

3. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 1, characterized in that: In step S2, each feature extraction unit includes an improved recurrent convolutional network block, namely an improved RLConvNeXt block and a similarity-based attention mechanism layer SimAM. Improved RLConvNeXt block: extracts multi-dimensional channel features through convolutions and residual connections of different dimensions; Similarity-based attention mechanism layer SimAM: The multi-dimensional channel features of the input are used to capture features and contextual information at different levels and then output.

4. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 3, characterized in that: The improved RLConvNeXt block includes an enhanced local attention layer ELA, a deep convolutional layer DwConv, a LayerNorm layer, a convolutional layer 1 Conv1, a GELU layer, and a convolutional layer 2 Conv2. After the input features pass through the ELA layer, they are input into the deep convolutional layer DwConv to capture the spatial features of the image. After that, they pass through the LayerNorm layer and the convolutional layer 1 Conv1, and then undergo a nonlinear transformation by the GELU layer. After entering the convolutional layer 2 Conv2, the convolved features are output and residually connected with the input features to obtain the output features.

5. The full-resolution network-based OCTA image retinal vessel segmentation method according to any one of claims 1 to 4, characterized in that: In step S3, the cascade feature enhancement module includes a cross-layer guidance module, a hybrid attention module and a multi-axis attention module. Cross-layer guidance module: Through skip connections, the image features of the 1st and 2nd layers (i.e., low-level detail features) output by the feature encoding module are fused with the image features of the 3rd to 5th layers (i.e., high-level semantic features) and then output to the hybrid attention module; Hybrid Attention Module: It includes five hybrid attention layers. Each hybrid attention layer highlights important features through spatial attention and channel attention mechanisms and outputs them to the multi-axis attention module. Multi-axis attention module: Five multi-axis attention blocks (GMA Blocks) are used, and each mixed attention layer is input into each multi-axis attention block (GMA Block) one by one. Each multi-axis attention block (GMA Block) captures multi-scale features from multiple spatial dimensions, and the output features of the five multi-axis attention blocks (GMA Blocks) are vector-added and output.

6. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 5, characterized in that: The cross-layer guidance module includes a first jump connection layer GAB1, a second jump connection layer GAB2, a third jump connection layer GAB3 and a fourth jump connection layer GAB4. The first skip connection layer GAB1: fuses the first-level image features output by the first feature extraction unit with the second-level image features output by the second feature extraction unit through a skip connection and outputs a fused feature 1; The second jump connection layer GAB2: through the jump connection, the third-level image features output by the third feature extraction unit are fused with the fusion feature 1 and the fusion feature 2 is output; The third skip connection layer GAB2: fuses the fourth-level image features output by the fourth feature extraction unit with the fusion feature two through the skip connection and outputs the fusion feature three; The fourth jump connection layer GAB4: through the jump connection, the fifth-level image feature output by the fifth feature extraction unit is fused with the fusion feature three and the fusion feature four is output.

7. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 6, characterized in that: In the hybrid attention module, the first hybrid attention layer Hybrid Attn1 highlights the important features of the first level of input image features and outputs them; the second hybrid attention layer Hybrid Attn2 highlights the important features of the input fusion feature one and outputs them; the third hybrid attention layer Hybrid Attn3: highlights the important features of the input fusion feature two and outputs them; the fourth hybrid attention layer Hybrid Attn4 highlights the important features of the input fusion feature three and outputs them; The fifth hybrid attention layer Hybrid Attn5 outputs the fused features of the input and highlights the important features.

8. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 7, characterized in that: The hybrid attention layer includes a channel attention mechanism module, a spatial attention mechanism module, and a residual connection layer. In the channel attention mechanism module, the input image features pass through the average pooling layer 1 Avg Pool1 and are output to Reshape1 and average pooling layer 2 Avg Pool2 respectively. The output features of the reshaping layer 1 Reshape1 pass through the convolution layer 1 Conv1 and are output to the reshaping layer 2 Reshape2. The output features of the average pooling layer 2 Avg Pool2 pass through the reshaping layer 3 Reshape3, the convolution layer 2 Conv2, and the average pooling layer 3 Avg Pool3 in sequence. The output features of the reshaping layer 2 Reshape2 are residually connected with the output features of the average pooling layer 3 Avg Pool3 and then output to the average pooling layer 4 Avg Pool4. The output features of the average pooling layer 4 AvgPool4 are weighted with the input image features channel by channel and then output. In the spatial attention mechanism module, the input image features are respectively average pooled along the X-axis direction by the average pooling layer X Avg Pool and average pooled along the Y-axis direction by the average pooling layer Y Avg Pool. The output features of the average pooling layer XAvg Pool are sequentially passed through the convolution layer three Conv3, the group normalization layer GroupNorm1 and the activation function layer Sigmoid1 to output the X-direction spatial weight map. The output features of the average pooling layer Y Avg Pool are sequentially passed through the convolution layer four Conv4, the group normalization layer GroupNorm2 and the activation function layer Sigmoid2 to output the Y-direction spatial weight map. The X-direction spatial weight map and the Y-direction spatial weight map are multiplied point by point and then multiplied with the input image features for output. Residual connection layer: The output features of the channel attention mechanism module are residually connected with the output features of the spatial attention mechanism module and then output.

9. The full-resolution network-based OCTA image retinal vessel segmentation method according to claim 5, characterized in that: The multi-axis attention block GMA Block includes LayerNorm layer 1, two-dimensional separable convolution layer, one-dimensional convolution layer 1, one-dimensional convolution layer 2, depth-wise separable convolution layer, splicing layer and LayerNorm layer 2. The input image feature map is normalized by LayerNorm layer 1 and then output to the two-dimensional separable convolution layer, one-dimensional convolution layer 1, one-dimensional convolution layer 2, and depth-wise separable convolution layer; 2D separable convolution layer: As the XY branch, it uses 2D separable convolution to process the original spatial structure of the input feature map, models the spatial attention on the XY axis plane, and outputs it to the splicing layer; One-dimensional convolution layer 1: As the ZX branch, it transposes the input feature map in the width dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; One-dimensional convolution layer 2: As the ZY branch, it transposes the input feature map in the height dimension, uses one-dimensional convolution to model spatial attention, and then outputs it to the splicing layer; Depthwise separable convolutional layer: extracts local spatial context information from the input feature map and outputs it to the splicing layer; Splicing layer: Splicing is performed on the channel dimension and then output to the LayerNorm layer 2; LayerNorm layer 2: After normalization, it outputs a feature map that integrates multi-dimensional spatial attention information.

10. The full-resolution network-based OCTA image retinal vessel segmentation method according to any one of claims 1 to 4, characterized in that: In step S5, the loss is calculated using the DiceJaccardLoss loss function, which is a weighted combination of the Dice loss and the Jaccard loss.

Images