A retinal blood vessel image segmentation method based on direction selection and scale relationship calibration

Abstract

The application discloses a retinal blood vessel image segmentation method based on direction selection and scale relationship calibration, comprising the following steps: acquiring a color fundus image containing retinal blood vessels, and preprocessing to obtain a network input image; inputting the network input image into an encoding network; the encoding network sequentially performs first, second, third and fourth encoding stages to obtain four scale encoding features; the four scale encoding features are respectively subjected to scale relationship calibration to obtain four scale calibration features corresponding to the four scale encoding features; the four scale calibration features are input into a decoder; the decoder outputs final decoding features; and the final decoding features are subjected to segmentation prediction to obtain a retinal blood vessel segmentation result. The application solves the problems of insufficient attention to different direction blood vessel structures, simple fusion of different scale features leading to noise interference and detail loss, and insufficient recovery of blood vessel edge and end structure in the decoding upsampling process in the existing retinal blood vessel segmentation method.

Description

Technical Field

[0001] This invention relates to medical image processing technology, specifically to a retinal vessel image segmentation method based on orientation selection and scale relationship calibration. Background Technology

[0002] Retinal vessels are important structures in fundus images. Accurate segmentation of retinal vessels helps to obtain their morphology, branching, and connectivity features, facilitating disease screening. However, fundus images are easily affected by factors such as illumination, noise, and background texture. Small vessels and their terminals have little difference from the surrounding background, making segmentation prone to problems such as missed vessels, breaks, and unclear boundaries.

[0003] Existing methods for retinal vessel segmentation mainly include traditional image processing methods and deep learning methods. Traditional image processing methods often rely on manually designed filters, thresholds, or edge features, making them sensitive to image quality and parameter settings, and difficult to stably adapt to different imaging conditions. While deep learning-based methods can automatically extract vessel features, conventional encoder-decoder networks, such as UNet, typically rely on fixed convolutions and simple skip fusion, paying insufficient attention to the directional structure of vessels with different orientations. This can easily weaken small vessels during the encoding process, leading to segmentation results with vessel breaks or missed branches. Furthermore, existing networks often use concatenation or addition to fuse features at different scales. Background textures and noise from shallow features are easily carried into the decoding process, while deep features may weaken vessel edge and terminal details due to reduced resolution, resulting in incomplete segmentation of low-contrast vessel regions and unclear boundaries. All these problems significantly affect the accuracy of retinal vessel segmentation results. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a retinal vessel image segmentation method based on orientation selection and scale relationship calibration, which can reduce background noise interference and loss of vascular details caused by simple fusion of features at different scales, and enhance the complementarity between vascular features at different scales.

[0005] Technical solution: The present invention provides a retinal vessel image segmentation method based on orientation selection and scale relationship calibration, comprising:

[0006] S1. Obtain a color fundus image containing retinal vessels, and preprocess the color fundus image to obtain a network input image;

[0007] S2. Input the network input image obtained in step S1 into the encoding network. The encoding network sequentially executes the first encoding stage, the second encoding stage, the third encoding stage, and the fourth encoding stage to obtain four scale encoding features.

[0008] S3. Perform scale relationship calibration on the four scale-coded features obtained in step S2 to obtain four scale-calibrated features corresponding to the four scale-coded features. The scale relationship calibration includes four processes: global scale description, scale relationship modeling, channel calibration coefficient generation, and channel recalibration. The global scale description is used to extract the overall channel response of each scale-coded feature. The scale relationship modeling is used to calculate the correlation between different scale-coded features. The channel calibration coefficient generation is used to convert the correlation into channel weights corresponding to each scale. The channel recalibration is used to enhance or suppress the channel response of each scale-coded feature using the channel weights.

[0009] S4. Input the four scale calibration features obtained in step S3 into the decoder, and the decoder outputs the final decoded features. The decoding process of the decoder includes three processes: third-scale decoding, second-scale decoding, and first-scale decoding. The third-scale decoding is used to upsample the fourth-scale calibration feature and fuse it with the third-scale calibration feature. The second-scale decoding is used to upsample the third-scale decoding result and fuse it with the second-scale calibration feature. The first-scale decoding is used to upsample the second-scale decoding result and fuse it with the first-scale calibration feature.

[0010] S5. Perform segmentation prediction on the final decoded features obtained in step S4 to obtain the retinal vessel segmentation result; the segmentation prediction includes 1×1 convolution mapping, Sigmoid mapping and threshold binarization processing, wherein the 1×1 convolution mapping is used to convert the final decoded features into single-channel prediction features, the Sigmoid mapping is used to perform nonlinear mapping on the single-channel prediction features to obtain a retinal vessel probability map, and the threshold binarization processing is used to generate a retinal vessel binary segmentation map based on the retinal vessel probability map.

[0011] Furthermore, in step S2, the input image obtained in step S1 is input into the encoding network, and the encoding network sequentially executes the first encoding stage, the second encoding stage, the third encoding stage, and the fourth encoding stage to obtain four scale-coded features, including:

[0012] The first encoding stage involves performing 3×3 convolution, batch normalization, and ReLU activation on the network input image to obtain the initial features of the first stage. Three directional feature extractions are then performed on these initial features, executed sequentially with the output of the previous extraction serving as the input for the next, thus obtaining the output features of the first stage. Simultaneously, the network input image is also subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the first input branch features. Finally, the output features of the first stage are concatenated with the first input branch features along the channel dimension to obtain the first scale encoded features.

[0013] The second encoding stage involves bilinear interpolation downsampling the network input image to adjust its spatial size to half that of the original network input image, resulting in a half-scale input image. The first-scale encoded features are downsampled and then subjected to batch normalization and ReLU activation to obtain the initial features for the second stage. Three consecutive directional feature extractions are performed on these initial features to obtain the second-stage output features. Simultaneously, the half-scale input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the second input branch features. The second-stage output features are then concatenated with the second input branch features along the channel dimension to obtain the second-scale encoded features.

[0014] The third encoding stage involves bilinear interpolation downsampling the network input image to adjust its spatial size to one-quarter of the original network input image, resulting in a one-quarter scale input image. The second scale encoded features are downsampled and then processed with batch normalization and ReLU activation to obtain the initial features for the third stage. Nine consecutive directional feature extractions are performed on these initial features to obtain the third stage output features. Simultaneously, the one-quarter scale input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the third input branch features. Finally, the third stage output features and the third input branch features are concatenated along the channel dimension to obtain the third scale encoded features.

[0015] Fourth encoding stage: The third-scale encoded features are downsampled and processed by batch normalization and ReLU activation to obtain the initial features of the fourth stage; the initial features of the fourth stage are subjected to three consecutive directional feature extractions to obtain the output features of the fourth stage; the output features of the fourth stage are used as the fourth-scale encoded features.

[0016] Furthermore, the process of extracting the directional features is as follows:

[0017] The input image to be extracted from the network is denoted as X. First, X is subjected to 1×1 convolution, batch normalization and ReLU activation to obtain channel extended features.

[0018] The channel expansion features are input into the horizontal depth convolution branch, the vertical depth convolution branch, the local 2D depth convolution branch, and the large receptive field depth convolution branch, respectively, to obtain horizontal features, vertical features, local structural features, and large receptive field context features. Specifically, the horizontal depth convolution branch uses a 1×7 depth convolution to extract the continuous response of horizontally extending blood vessels; the vertical depth convolution branch uses a 7×1 depth convolution to extract the continuous response of vertically extending blood vessels; the local 2D depth convolution branch uses a 7×7 depth convolution to extract tortuous blood vessels, bifurcation regions, and local intersection structures; and the large receptive field depth convolution branch uses a 7×7 depth convolution with an dilation rate of 2 to supplement vascular context information over a larger area.

[0019] The lateral direction features, vertical direction features, local structural features, and large receptive field context features are all processed by batch normalization and ReLU activation.

[0020] Based on the channel expansion features, selection weights are generated for the four depthwise convolutional branches. These selection weights are then used to weight and fuse the lateral features, vertical features, local structural features, and large receptive field context features to obtain directional fusion features. The weighted fusion process can be represented as follows:

[0021] ;

[0022] in, Indicates directional fusion features, Indicates lateral characteristics. Indicates vertical direction features. Indicates local structural features, Indicates the contextual features of a large receptive field; , , and These represent the selection weights corresponding to the four depthwise convolutional branches;

[0023] The directional fusion features are recalibrated by channels and projected back to the channel dimension of the input feature map X using a 1×1 convolution. After batch normalization and layer scaling, they are added to the input feature map X through residual connections to obtain the output features of directional feature extraction. The layer scaling process is to scale the feature response after pointwise convolution projection using a learnable scaling factor.

[0024] Furthermore, in step S3, the scale relationship calibration of the four scale-coded features obtained in step S2 is performed respectively to obtain four scale-calibrated features corresponding to the four scale-coded features, including:

[0025] Global scale description: Global average pooling is performed on the four scale-coded features to obtain four scale description vectors corresponding to the four scale-coded features; the scale description vectors are used to represent the overall response of the corresponding scale-coded features in the channel dimension.

[0026] Scale relationship modeling: Taking four scale description vectors as input, the scale relationship matrix A is obtained through query vector and key vector generation, scale similarity calculation and normalization. The scale relationship matrix A is used to represent the degree of reference of each scale coding feature to other scale coding features when calibrating.

[0027] Channel calibration coefficient generation: Learnable linear mapping is performed on the four scale description vectors to obtain four scale representation vectors of the same dimension; the four scale representation vectors are weighted according to the scale relationship matrix A to obtain the cross-scale relationship description vectors corresponding to each scale; then each cross-scale relationship description vector is mapped to the channel dimension of the corresponding scale encoded feature to generate the channel calibration coefficients corresponding to the four scales.

[0028] Channel recalibration: Using four channel calibration coefficients, the channel responses of the four scale-coded features are adjusted respectively to obtain four scale-calibrated features; the channel recalibration process is expressed as follows:

[0029] ;

[0030] in, Indicates the first Each scale encodes features; Indicates the first Channel calibration coefficients corresponding to each scale; Indicates the first Individual scale calibration features; This indicates element-wise multiplication.

[0031] Furthermore, taking four scale description vectors as input, and through query vector and key vector generation, scale similarity calculation, and normalization processing, the scale relationship matrix A is obtained, including:

[0032] Query vector and key vector generation: Let the first... The scale description vectors are , No. The channel dimension of each scale description vector is ,Will Multiply by the query mapping weight and the key mapping weight respectively to obtain the first... The query vector corresponding to each scale and the The key vector corresponding to each scale :

[0033] ;

[0034] in, Indicates the first Each scale describes a vector; Indicates the first The query vector corresponding to each scale; Indicates the first The key vector corresponding to each scale; Indicates the first Each scale corresponds to a query mapping weight; Indicates the first The key mapping weights corresponding to each scale; Indicates the first Each scale corresponds to a query mapping bias term; Indicates the first Each scale corresponds to a key mapping bias term;

[0035] Inter-scale similarity calculation: The first scale... The query vector corresponding to each scale With the The key vector corresponding to each scale Perform a dot product operation and normalize using a dimension factor to obtain the ... The first scale and the first Similarity scores between each scale :

[0036] ;

[0037] in, Indicates the first The first scale and the first Similarity scores between each scale; This represents a unified dimension for both the query vector and the key vector. Used for scaling the dot product result;

[0038] Normalization processing: for the first The similarity scores between each scale are then Softmax normalized to obtain the i-th scale. The scale relation weights corresponding to the i-th scale; for the i-th scale... The first scale and the first The scale, the scale relation matrix A of the nth scale. Line number The elements of a column are represented as follows:

[0039] ;

[0040] in, Indicates the first The scale-coded features are used for calibration during the first-level calibration. Reference weights for each scale-coded feature; Indicates the first The first scale and the first Similarity scores between each scale; Indicates the first The first scale and the first Similarity scores between each scale;

[0041] Through the above normalization process, the first The sum of the four association weights corresponding to each scale is 1, so that the similarity scores between different scales are transformed into a weight distribution for feature calibration. Composition of scale relation matrix A:

[0042] ;

[0043] in, This represents a query matrix composed of query vectors corresponding to four scales; This represents a key matrix composed of key vectors corresponding to four scales; This represents the unified dimension of the query vector and the key vector.

[0044] Furthermore, in step S4, the four scale calibration features obtained in step S3 are input into the decoder, and the decoder outputs the final decoded features, including:

[0045] Third-scale decoding: Upsample the fourth-scale calibration features so that the spatial dimensions of the upsampled fourth-scale calibration features are consistent with those of the third-scale calibration features; concatenate the upsampled features and the third-scale calibration features along the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the third-scale fused features; guide the refinement and fusion of the third-scale fused features to obtain the third-scale decoded features.

[0046] Second-scale decoding: Upsample the third-scale decoding features so that the spatial dimensions of the upsampled third-scale decoding features are consistent with the second-scale calibration features; concatenate the upsampled features and the second-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the second-scale fusion features; guide the refinement fusion of the second-scale fusion features to obtain the second-scale decoding features.

[0047] First-scale decoding: Upsample the second-scale decoding features so that the spatial dimensions of the upsampled second-scale decoding features are consistent with the first-scale calibration features; concatenate the upsampled features and the first-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the first-scale fusion features; guide the refinement fusion of the first-scale fusion features to obtain the first-scale decoding features, and use the first-scale decoding features as the final decoding features.

[0048] Furthermore, the specific process of the rapid guided refinement fusion is as follows:

[0049] The current scale fusion features are used as guiding features, and the features before upsampling in this decoding stage are used as features to be refined. The linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined in the local window. The current scale fusion features are then linearly reconstructed using the linear transformation coefficients and bias terms to obtain the decoding features of the current decoding stage.

[0050] Furthermore, the linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined within the local window; then, the linear transformation coefficients and bias terms are used to perform linear reconstruction of the current scale fused features to obtain the decoding features of the current decoding stage, including:

[0051] Using bilinear interpolation downsampling, the height and width of the guiding feature are adjusted to match the feature to be refined, resulting in a size-adjusted guiding feature.

[0052] The response change relationship within a local window is statistically analyzed between the guided feature after size adjustment and the feature to be refined, and the linear transformation coefficients and bias terms are calculated based on the response change relationship; the linear transformation coefficients and bias terms are represented in the form of a feature map, and the height and width of the feature map are consistent with the feature to be refined;

[0053] Bilinear interpolation upsampling is performed on the feature maps representing linear transform coefficients and bias terms so that the height and width of the upsampled feature maps of linear transform coefficients and bias terms are consistent with the guiding features;

[0054] Linear reconstruction of the current scale-fused features is performed using the upsampled linear transform coefficients and bias term to obtain the decoding features for the current decoding stage:

[0055] ;

[0056] in, Indicates the guiding features at the current scale; This indicates that the height and width have been adjusted to match. Consistent linear transformation coefficients; This indicates that the height and width have been adjusted to match. Consistent bias terms; This indicates the decoding characteristics at the current decoding stage; This indicates element-wise multiplication.

[0057] Furthermore, the specific process of the Sigmoid mapping is as follows:

[0058] A nonlinear mapping is applied to the single-channel predicted features obtained by the 1×1 convolution mapping to obtain a retinal vessel probability map; the value of each pixel in the retinal vessel probability map ranges from 0 to 1, representing the probability that the pixel belongs to a vascular region; the Sigmoid mapping is expressed as:

[0059] ;

[0060] in, This indicates the single-channel predicted feature at the pixel location. The response value at that location, Indicates pixel position The probability that it belongs to a vascular region.

[0061] Furthermore, the specific process of the threshold binarization is as follows:

[0062] The retinal vessel probability map is judged based on a preset threshold to obtain a binary segmentation map of the retinal vessels; the threshold binarization process is expressed as follows:

[0063] ;

[0064] in, Indicates pixel position The binary segmentation result at the location, Indicates pixel position The corresponding pixel belongs to the blood vessel region. Indicates pixel position The corresponding pixels belong to the background area; This represents a preset threshold; when the probability of a blood vessel corresponding to a pixel location is greater than or equal to the preset threshold, the pixel location... The corresponding pixel is determined to be a blood vessel region; when the probability of a blood vessel corresponding to a pixel position is less than a preset threshold, the pixel is determined to be a background region.

[0065] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows:

[0066] (1) In the process of directional feature extraction in step S2, vascular features in different directions and under different receptive fields are extracted by horizontal, vertical, local two-dimensional and large receptive field depth convolution branches, and weighted fusion is performed according to the selected weights to improve the problem of insufficient attention to vascular structures with different directions in existing methods and improve the feature expression ability of small blood vessels and complex vascular structures.

[0067] (2) In step S3, the overall channel response of each scale feature is obtained through global scale description, the correlation between different scales is established through scale relationship modeling, and channel calibration coefficients are generated for channel recalibration to reduce background noise interference and loss of vascular details caused by simple fusion of different scale features and enhance the complementarity between vascular features of different scales.

[0068] (3) In step S4, the low-resolution features are refined by using the structural information in the current scale fusion features through guided refinement fusion, which improves the problem of insufficient recovery of the blood vessel edge and terminal structure in the ordinary upsampling process and improves the continuity, integrity and boundary clarity of the retinal vessel segmentation results. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the present invention.

[0070] Figure 2 This is a flowchart of the directional feature extraction process in step S2;

[0071] Figure 3 This is a schematic diagram of the scale relationship calibration process in step S3;

[0072] Figure 4 This is a flowchart illustrating the process of quickly guiding the refined fusion in step S4.

[0073] Figure 5 This is a comparison of the retinal vessel segmentation results of the present invention with those of U-Net, Attention U-Net, and UNet++ on the DRIVE and STARE datasets. Detailed Implementation

[0074] The technical solution of the present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0075] like Figure 1 As shown, the present invention provides a retinal vessel image segmentation method based on orientation selection and scale relationship calibration, comprising the following steps:

[0076] S1. Obtain a color fundus image containing retinal vessels, and preprocess the color fundus image to obtain the network input image.

[0077] The specific implementation process of step S1 is as follows:

[0078] S1.1. Obtain a color fundus image, which includes a red channel, a green channel, and a blue channel.

[0079] S1.2 Adjust the size of the color fundus image to meet the input size requirements of the coding network, resulting in an adjusted image with size H×W×3.

[0080] S1.3. Convert the adjusted image into tensor form and normalize the pixel values ​​to obtain the network input image. The network input image can be represented as:

[0081] ;

[0082] In the formula, Indicates batch size; 3 indicates the number of channels in the color fundus image; and These represent the image height and image width, respectively.

[0083] By standardizing the size, performing tensor transformations, and normalizing pixel values, the differences in imaging size, brightness range, and channel response among different color fundus images can be reduced, providing a relatively consistent input for subsequent coding networks and improving the stability of the feature extraction process.

[0084] S2. Input the network input image obtained in step S1 into the encoding network. The encoding network sequentially executes the first encoding stage, the second encoding stage, the third encoding stage, and the fourth encoding stage to obtain four scale encoding features.

[0085] The specific implementation process of step S2 is as follows:

[0086] S2.1, First Encoding Stage: The network input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the initial features of the first stage; the initial features of the first stage are subjected to three directional feature extractions, which are performed sequentially, with the output of the previous directional feature extraction serving as the input of the next directional feature extraction, thus obtaining the output features of the first stage; simultaneously, the network input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the first input branch features; the output features of the first stage are concatenated with the first input branch features along the channel dimension to obtain the first scale encoded features;

[0087] The first input branch feature is used to preserve the shallow texture and edge information in the network input image. The first input branch feature is concatenated with the first stage output feature obtained after three directional feature extractions. This can enhance the directional response of blood vessels while supplementing the local details of small blood vessels and blood vessel edges.

[0088] In step S2.1, orientation feature extraction refers to the process of sequentially performing channel expansion, orientation and receptive field branch extraction, branch weighted fusion, channel recalibration, pointwise convolutional projection, and residual connection on the input feature map. Specifically, the feature map of the input orientation feature extraction process is denoted as X. First, X is subjected to 1×1 convolution, batch normalization, and ReLU activation to obtain the channel expansion features.

[0089] The channel expansion features are then input into the horizontal depth convolution branch, the vertical depth convolution branch, the local 2D depth convolution branch, and the large receptive field depth convolution branch, respectively, to obtain horizontal features, vertical features, local structural features, and large receptive field context features. Specifically, the horizontal depth convolution branch uses a 1×7 depth convolution to extract continuous responses from horizontally extending blood vessels; the vertical depth convolution branch uses a 7×1 depth convolution to extract continuous responses from vertically extending blood vessels; the local 2D depth convolution branch uses a 7×7 depth convolution to extract tortuous blood vessels, bifurcation regions, and local intersection structures; and the large receptive field depth convolution branch uses a 7×7 depth convolution with a dilation rate of 2 to supplement vascular context information over a larger area.

[0090] By extracting vascular responses in different directions and spatial ranges through the four depth convolution branches mentioned above, the bias of a single convolution shape on blood vessels with specific orientations can be reduced, allowing for more complete feature expression of tortuous blood vessels, intersecting regions, and fine branches.

[0091] Lateral features, vertical features, local structural features, and large receptive field context features were all processed by batch normalization and ReLU activation.

[0092] Based on the channel expansion features, selection weights are generated for the four depthwise convolutional branches. These selection weights are then used to weight and fuse the lateral features, vertical features, local structural features, and large receptive field context features to obtain the directional fusion features. The weighted fusion process can be represented as follows:

[0093] ;

[0094] in, Indicates directional fusion features, Indicates lateral characteristics. Indicates vertical direction features. Indicates local structural features, Indicates the contextual features of a large receptive field; , , and These represent the selection weights corresponding to the four depthwise convolutional branches.

[0095] Adaptively determining the selection weights of each depth convolution branch based on channel expansion features can enhance the branch response that matches the current blood vessel orientation and local structure, and suppress the weakly correlated branch response, thereby reducing the interference of background texture on directional feature extraction.

[0096] The directional fusion features are recalibrated by channels and projected back to the channel dimension of the input feature map X using a 1×1 convolution. After batch normalization and layer scaling, they are added to the input feature map X through residual connections to obtain the output features of directional feature extraction. The layer scaling process involves scaling the feature response after pointwise convolution projection using learnable scaling factors. The scaling factors are updated along with other network parameters during network training.

[0097] Channel recalibration is used to highlight channel responses related to vascular structures, and pointwise convolutional projection is used to recover input features. Figure 1 The consistent channel dimension, layer scaling, and residual connections help preserve the original features and stabilize the continuous directional feature extraction process, reducing the effective detail attenuation caused by deep stacking.

[0098] S2.2, Second Encoding Stage: The network input image is downsampled using bilinear interpolation to adjust its spatial size to half that of the original network input image, resulting in a half-scale input image. The first-scale encoding features are then downsampled using a 2×2 convolution with a stride of 2, followed by batch normalization and ReLU activation to obtain the initial features for the second stage. These initial features are then subjected to three consecutive directional feature extraction steps as described in S2.1 to obtain the second-stage output features. Simultaneously, the half-scale input image is subjected to a 3×3 convolution, batch normalization, and ReLU activation to obtain the second input branch features. The second-stage output features are then concatenated with the second input branch features along the channel dimension to obtain the second-scale encoding features.

[0099] S2.3, Third Encoding Stage: The network input image is downsampled using bilinear interpolation to adjust its spatial size to one-quarter of the original network input image, resulting in a one-quarter scale input image. The second scale encoding features are downsampled using a 2×2 convolution with a stride of 2, followed by batch normalization and ReLU activation to obtain the initial features for the third stage. These initial features are then subjected to nine consecutive directional feature extractions as described in step S2.1 to obtain the third stage output features. Simultaneously, the one-quarter scale input image is processed using a 3×3 convolution, batch normalization, and ReLU activation to obtain the third input branch features. The third stage output features and the third input branch features are concatenated along the channel dimension to obtain the third scale encoding features.

[0100] S2.4, Fourth Encoding Stage: The third-scale encoding features are downsampled using a 2×2 convolution with a stride of 2, and then processed by batch normalization and ReLU activation to obtain the initial features of the fourth stage; the initial features of the fourth stage are subjected to three consecutive directional feature extraction steps in S2.1 to obtain the output features of the fourth stage; the output features of the fourth stage are used as the fourth-scale encoding features.

[0101] The second and third encoding stages introduce half-scale and quarter-scale input images, respectively, which can supplement the original image information of the corresponding scale during the network's progressive downsampling process, reducing the loss of fine blood vessel texture and edge information caused by continuous downsampling; the fourth encoding stage further expands the feature perception range, which is conducive to forming high-level features that include the overall direction of blood vessels and contextual relationships.

[0102] S3. Perform scale relationship calibration on the four scale-coded features obtained in step S2 to obtain four scale-calibrated features corresponding to the four scale-coded features. Scale relationship calibration includes four processes: global scale description, scale relationship modeling, channel calibration coefficient generation, and channel recalibration. Among them, global scale description is used to extract the overall channel response of each scale-coded feature, scale relationship modeling is used to calculate the correlation between different scale-coded features, channel calibration coefficient generation is used to convert the correlation into channel weights corresponding to each scale, and channel recalibration is used to enhance or suppress the channel response of each scale-coded feature using the channel weights.

[0103] The specific implementation process of step S3 is as follows:

[0104] S3.1 Global Scale Description: Global average pooling is performed on the four scale-coded features to obtain four scale description vectors corresponding to the four scale-coded features; the scale description vectors are used to represent the overall response of the corresponding scale-coded features in the channel dimension;

[0105] S3.2 Scale Relationship Modeling: Using four scale description vectors as input, the scale relationship matrix A is obtained through query vector and key vector generation, scale similarity calculation and normalization. The scale relationship matrix A is used to represent the degree of reference of each scale coding feature to other scale coding features when calibrating.

[0106] The specific implementation process of step S3.2 is as follows:

[0107] S3.2.1 Query Vector and Key Vector Generation: A learnable linear mapping is performed on the description vectors at each scale. Let the first... The scale description vectors are , No. The channel dimension of each scale description vector is ,Will Multiply by the query mapping weight and the key mapping weight respectively to obtain the first... The query vector corresponding to each scale and the The key vector corresponding to each scale :

[0108] ;

[0109] in, Indicates the first Each scale describes a vector; Indicates the first The query vector corresponding to each scale; Indicates the first The key vector corresponding to each scale; Indicates the first Each scale corresponds to a query mapping weight; Indicates the first The key mapping weights corresponding to each scale; Indicates the first Each scale corresponds to a query mapping bias term; Indicates the first The key mapping bias terms correspond to each scale. The aforementioned mapping weights and bias terms are learnable parameters in the model, updated along with the network parameters during training. and Used to the first Each scale description vector is derived from the original channel dimension. Mapping to a unified dimension This enables similarity calculations to be performed on descriptive vectors of different scales within the same dimensional space.

[0110] S3.2.2, Calculation of inter-scale similarity: The first... The query vector corresponding to each scale With the The key vector corresponding to each scale Perform a dot product operation and normalize using a dimension factor to obtain the ... The first scale and the first Similarity scores between each scale :

[0111] ;

[0112] in, Indicates the first The first scale and the first Similarity scores between each scale; This represents a unified dimension for both the query vector and the key vector. Used to scale the dot product result.

[0113] S3.2.3, Normalization processing: For the first... The similarity scores between each scale are then Softmax normalized to obtain the i-th scale. The scale relation weights corresponding to the i-th scale; for the i-th scale... The first scale and the first The scale, the scale relation matrix A of the nth scale. Line number The elements of a column can be represented as:

[0114] ;

[0115] in, Indicates the first The scale-coded features are used for calibration during the first-level calibration. Reference weights for each scale-coded feature; Indicates the first The first scale and the first Similarity scores between each scale; Indicates the first The first scale and the first Similarity scores between each scale;

[0116] Through the above normalization process, the first The sum of the four association weights corresponding to each scale is 1, so that the similarity scores between different scales are transformed into a weight distribution that can be used for feature calibration. Composition scale relation matrix :

[0117] ;

[0118] in, This represents a query matrix composed of query vectors corresponding to four scales; This represents a key matrix composed of key vectors corresponding to four scales; This represents the unified dimension of the query vector and the key vector.

[0119] After Softmax normalization, the reference degree of each scale to other scales is converted into a weight distribution, which enables scale relationship modeling to adaptively determine the contribution of information at different scales based on the content of the input image, thus alleviating the problem of insufficient adaptability of fixed fusion ratio to blood vessels of different sizes.

[0120] S3.3, Channel calibration coefficient generation: Perform learnable linear mapping on the four scale description vectors to obtain four scale representation vectors with a unified dimension; perform relational weighting on the four scale representation vectors according to the scale relation matrix A to obtain the cross-scale relation description vectors corresponding to each scale; then map each cross-scale relation description vector to the channel dimension of the corresponding scale encoded feature to generate the channel calibration coefficients corresponding to the four scales.

[0121] S3.4 Channel Recalibration: Using four channel calibration coefficients, the channel responses of the four scale-coded features are adjusted respectively, resulting in four scale-calibrated features. Thus, each scale-coded feature maintains its original spatial dimensions while adjusting its channel response. This preserves information about vessel edges and fine branches in superficial features while utilizing the overall structural information provided by deeper features, offering complementary multi-scale features for progressive decoding. The channel recalibration process can be represented as:

[0122] ;

[0123] in, Indicates the first Each scale encodes features; Indicates the first Channel calibration coefficients corresponding to each scale; Indicates the first Individual scale calibration features; This indicates element-wise multiplication.

[0124] S4. Input the four scale calibration features obtained in step S3 into the decoder, and the decoder outputs the final decoded features. The decoding process of the decoder includes three processes: third-scale decoding, second-scale decoding, and first-scale decoding. The third-scale decoding is used to upsample the fourth-scale calibration feature and fuse it with the third-scale calibration feature. The second-scale decoding is used to upsample the third-scale decoding result and fuse it with the second-scale calibration feature. The first-scale decoding is used to upsample the second-scale decoding result and fuse it with the first-scale calibration feature.

[0125] The above-mentioned step-by-step decoding method gradually restores high-level features to a higher spatial resolution, and introduces corresponding scale calibration features at each decoding stage, thereby reducing the spatial positioning deviation and structural detail loss caused by simple upsampling.

[0126] The specific implementation process of step S4 is as follows:

[0127] S4.1 Third-scale decoding: Upsample the fourth-scale calibration features so that the spatial dimensions of the upsampled fourth-scale calibration features are consistent with those of the third-scale calibration features; concatenate the upsampled features with the third-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the third-scale fused features; guide the refinement and fusion of the third-scale fused features to obtain the third-scale decoded features.

[0128] S4.2 Second-scale decoding: Upsample the third-scale decoding features so that the spatial dimensions of the upsampled third-scale decoding features are consistent with the second-scale calibration features; concatenate the upsampled features and the second-scale calibration features along the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the second-scale fused features; guide the refinement fusion of the second-scale fused features to obtain the second-scale decoding features;

[0129] S4.3 First-scale decoding: Upsample the second-scale decoding features so that the spatial dimensions of the upsampled second-scale decoding features are consistent with the first-scale calibration features; concatenate the upsampled features and the first-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the first-scale fused features; guide the refinement fusion of the first-scale fused features to obtain the first-scale decoding features, and use the first-scale decoding features as the final decoding features.

[0130] In steps S4.1 to S4.3, rapid guided refinement fusion refers to refining the low-resolution decoded features from the previous scale using high-resolution structural information from the current scale's fused features in each decoding stage. This ensures that the fused features retain both semantic information and vessel edge details, which is beneficial for recovering small vessels, vessel ends, and boundary transition regions that are difficult to fully reconstruct using conventional interpolation and upsampling. The specific process of rapid guided refinement fusion is as follows:

[0131] The current scale fusion features are used as guiding features, and the features before upsampling in this decoding stage are used as features to be refined. The linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined in the local window. Then, the current scale fusion features are linearly reconstructed using the linear transformation coefficients and bias terms to obtain the decoding features of the current decoding stage.

[0132] In this embodiment, linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined within a local window; then, the linear transformation coefficients and bias terms are used to perform linear reconstruction of the current scale fusion features to obtain the decoding features of the current decoding stage, including:

[0133] First, bilinear interpolation downsampling is used to adjust the height and width of the guiding feature to match the feature to be refined, thus obtaining the size-adjusted guiding feature;

[0134] Then, the response change relationship within the local window is statistically analyzed between the guided feature after size adjustment and the feature to be refined, and the linear transformation coefficients and bias terms are calculated based on the response change relationship; the linear transformation coefficients and bias terms are represented in the form of a feature map, and the height and width of the feature map are consistent with the feature to be refined;

[0135] Then, bilinear interpolation upsampling is performed on the feature maps representing the linear transformation coefficients and bias terms, so that the height and width of the upsampled feature maps of the linear transformation coefficients and bias terms are consistent with the guiding features;

[0136] Finally, the upsampled linear transform coefficients and bias term are used to perform linear reconstruction of the current scale fused features to obtain the decoding features for the current decoding stage. The decoding features obtained through fast guided refinement fusion can be expressed as:

[0137] ;

[0138] in, Indicates the guiding features at the current scale; This indicates that the height and width have been adjusted to match. Consistent linear transformation coefficients; This indicates that the height and width have been adjusted to match. Consistent bias terms; This indicates the decoding characteristics at the current decoding stage; This indicates element-wise multiplication.

[0139] Since the linear transformation coefficients and bias terms are first calculated in the low-resolution space and then restored to the spatial size of the guiding features, the computational cost required to directly calculate local relationships on high-resolution features can be reduced while refining the vascular structure using local response relationships.

[0140] S5. Perform segmentation prediction on the final decoded features obtained in step S4 to obtain the retinal vessel segmentation result. The segmentation prediction includes 1×1 convolution mapping, Sigmoid mapping and threshold binarization processing. The 1×1 convolution mapping is used to convert the final decoded features into single-channel prediction features. The Sigmoid mapping is used to perform nonlinear mapping on the single-channel prediction features to obtain a retinal vessel probability map. The threshold binarization processing is used to generate a binary segmentation map of retinal vessels based on the retinal vessel probability map.

[0141] The specific implementation process of step S5 is as follows:

[0142] S5.1, 1×1 Convolution Mapping: The final decoded features are processed by 1×1 convolution, mapping their channel number to 1 to obtain single-channel prediction features. 1×1 convolution integrates the channel information of the final decoded features without changing the feature map space size, resulting in a blood vessel prediction response for each pixel location.

[0143] S5.2, Sigmoid Mapping: A non-linear mapping is applied to the single-channel predicted features obtained from the 1×1 convolution mapping to obtain a retinal vessel probability map. Each pixel in the retinal vessel probability map has a value ranging from 0 to 1, representing the probability that the pixel belongs to a vascular region. The Sigmoid mapping can be expressed as:

[0144] ;

[0145] in, This indicates the single-channel predicted feature at the pixel location. The response value at that location, Indicates pixel position The probability that it belongs to a vascular region.

[0146] S5.3 Threshold Binarization: The retinal vessel probability map is judged according to a preset threshold to obtain a binary segmentation map of the retinal vessels. Threshold binarization can be expressed as:

[0147] ;

[0148] in, Indicates pixel position The binary segmentation result at the location, Indicates pixel position The corresponding pixels belong to the blood vessel region. Indicates pixel position The corresponding pixels belong to the background area; This represents a preset threshold; in this embodiment, Set the value to 0.5; when the probability of a blood vessel corresponding to a pixel location is greater than or equal to 0.5, set the pixel location... The corresponding pixel is identified as a blood vessel region; when the probability of a blood vessel corresponding to a pixel location is less than 0.5, the pixel location is... The corresponding pixels are identified as the background area.

[0149] S5.4. Use the binary segmentation map of retinal vessels as the final output to represent the retinal vessel region in the color fundus image.

[0150] This invention can improve the continuity, integrity, and boundary clarity of retinal vessel segmentation results.

Claims

1. A retinal vessel image segmentation method based on orientation selection and scale relationship calibration, characterized in that, include: S1. Obtain a color fundus image containing retinal vessels, and preprocess the color fundus image to obtain a network input image; S2. Input the network input image obtained in step S1 into the encoding network. The encoding network sequentially executes the first encoding stage, the second encoding stage, the third encoding stage, and the fourth encoding stage to obtain four scale encoding features. S3. Perform scale relationship calibration on the four scale coding features obtained in step S2 to obtain four scale calibration features corresponding to the four scale coding features. The scale relationship calibration includes four processes: global scale description, scale relationship modeling, channel calibration coefficient generation, and channel recalibration. The global scale description is used to extract the overall channel response of the coding features at each scale. The scale relationship modeling is used to calculate the correlation between coding features at different scales. The channel calibration coefficient generation is used to convert the correlation into channel weights corresponding to each scale. The channel recalibration is used to enhance or suppress the channel response of the coding features at each scale using the channel weights. S4. Input the four scale calibration features obtained in step S3 into the decoder, and the decoder outputs the final decoded features. The decoding process of the decoder includes three processes: third-scale decoding, second-scale decoding, and first-scale decoding. The third-scale decoding is used to upsample the fourth-scale calibration feature and fuse it with the third-scale calibration feature. The second-scale decoding is used to upsample the third-scale decoding result and fuse it with the second-scale calibration feature. The first-scale decoding is used to upsample the second-scale decoding result and fuse it with the first-scale calibration feature. S5. Perform segmentation prediction on the final decoded features obtained in step S4 to obtain the retinal vessel segmentation result; the segmentation prediction includes 1×1 convolution mapping, Sigmoid mapping and threshold binarization processing, wherein the 1×1 convolution mapping is used to convert the final decoded features into single-channel prediction features, the Sigmoid mapping is used to perform nonlinear mapping on the single-channel prediction features to obtain a retinal vessel probability map, and the threshold binarization processing is used to generate a retinal vessel binary segmentation map based on the retinal vessel probability map.

2. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 1, characterized in that, In step S2, the input image obtained in step S1 is input into the encoding network. The encoding network sequentially executes the first encoding stage, the second encoding stage, the third encoding stage, and the fourth encoding stage to obtain four scale-coded features, including: The first encoding stage involves performing 3×3 convolution, batch normalization, and ReLU activation on the network input image to obtain the initial features of the first stage. Three directional feature extractions are then performed on these initial features, executed sequentially with the output of the previous extraction serving as the input for the next, thus obtaining the output features of the first stage. Simultaneously, the network input image is also subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the first input branch features. Finally, the output features of the first stage are concatenated with the first input branch features along the channel dimension to obtain the first scale encoded features. The second encoding stage involves bilinear interpolation downsampling the network input image to adjust its spatial size to half that of the original network input image, resulting in a half-scale input image. The first-scale encoded features are downsampled and then subjected to batch normalization and ReLU activation to obtain the initial features for the second stage. Three consecutive directional feature extractions are performed on these initial features to obtain the second-stage output features. Simultaneously, the half-scale input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the second input branch features. The second-stage output features are then concatenated with the second input branch features along the channel dimension to obtain the second-scale encoded features. The third encoding stage involves bilinear interpolation downsampling the network input image to adjust its spatial size to one-quarter of the original network input image, resulting in a one-quarter scale input image. The second scale encoded features are downsampled and then processed with batch normalization and ReLU activation to obtain the initial features for the third stage. Nine consecutive directional feature extractions are performed on these initial features to obtain the third stage output features. Simultaneously, the one-quarter scale input image is subjected to 3×3 convolution, batch normalization, and ReLU activation to obtain the third input branch features. Finally, the third stage output features and the third input branch features are concatenated along the channel dimension to obtain the third scale encoded features. Fourth encoding stage: The third-scale encoded features are downsampled and processed by batch normalization and ReLU activation to obtain the initial features of the fourth stage; the initial features of the fourth stage are subjected to three consecutive directional feature extractions to obtain the output features of the fourth stage; the output features of the fourth stage are used as the fourth-scale encoded features.

3. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 2, characterized in that, The process of extracting the directional features is as follows: The input feature map is denoted as X. First, X is subjected to 1×1 convolution, batch normalization and ReLU activation to obtain channel extended features. The channel expansion features are input into the horizontal depth convolution branch, the vertical depth convolution branch, the local 2D depth convolution branch, and the large receptive field depth convolution branch, respectively, to obtain horizontal features, vertical features, local structural features, and large receptive field context features. Specifically, the horizontal depth convolution branch uses a 1×7 depth convolution to extract the continuous response of horizontally extending blood vessels; the vertical depth convolution branch uses a 7×1 depth convolution to extract the continuous response of vertically extending blood vessels; the local 2D depth convolution branch uses a 7×7 depth convolution to extract tortuous blood vessels, bifurcation regions, and local intersection structures; and the large receptive field depth convolution branch uses a 7×7 depth convolution with an dilation rate of 2 to supplement vascular context information over a larger area. The lateral direction features, vertical direction features, local structural features, and large receptive field context features are all processed by batch normalization and ReLU activation. Based on the channel expansion features, selection weights are generated for the four depthwise convolutional branches. These selection weights are then used to weight and fuse the lateral features, vertical features, local structural features, and large receptive field context features to obtain directional fusion features. The weighted fusion process can be represented as follows: ； in, Indicates directional fusion features, Indicates lateral characteristics. Indicates vertical direction features. Indicates local structural features, Indicates the contextual features of a large receptive field; , , and These represent the selection weights corresponding to the four depthwise convolutional branches; The directional fusion features are recalibrated by channels and projected back to the channel dimension of the input feature map X using a 1×1 convolution. After batch normalization and layer scaling, they are added to the input feature map X through residual connections to obtain the output features of directional feature extraction. The layer scaling process is to scale the feature response after pointwise convolution projection using a learnable scaling factor.

4. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 1, characterized in that, In step S3, the scale relationship calibration of the four scale-coded features obtained in step S2 is performed to obtain four scale-calibrated features corresponding to the four scale-coded features, including: Global scale description: Global average pooling is performed on the four scale-coded features to obtain four scale description vectors corresponding to the four scale-coded features; the scale description vectors are used to represent the overall response of the corresponding scale-coded features in the channel dimension. Scale relationship modeling: Taking four scale description vectors as input, the scale relationship matrix A is obtained through query vector and key vector generation, scale similarity calculation and normalization. The scale relationship matrix A is used to represent the degree of reference of each scale coding feature to other scale coding features when calibrating. Channel calibration coefficient generation: Learnable linear mapping is performed on the four scale description vectors to obtain four scale representation vectors of the same dimension; the four scale representation vectors are weighted according to the scale relationship matrix A to obtain the cross-scale relationship description vectors corresponding to each scale; then each cross-scale relationship description vector is mapped to the channel dimension of the corresponding scale encoded feature to generate the channel calibration coefficients corresponding to the four scales. Channel recalibration: Using four channel calibration coefficients, the channel responses of the four scale-coded features are adjusted respectively to obtain four scale-calibrated features; the channel recalibration process is expressed as follows: ； in, Indicates the first Each scale encodes features; Indicates the first Channel calibration coefficients corresponding to each scale; Indicates the first Individual scale calibration features; This indicates element-wise multiplication.

5. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 4, characterized in that, The process involves taking four scale description vectors as input, generating query vectors and key vectors, calculating inter-scale similarity, and performing normalization to obtain a scale relationship matrix A, including: Query vector and key vector generation: Let the first... The scale description vectors are , No. The channel dimension of each scale description vector is ,Will Multiply by the query mapping weight and the key mapping weight respectively to obtain the first... The query vector corresponding to each scale and the The key vector corresponding to each scale : ； in, Indicates the first Each scale describes a vector; Indicates the first The query vector corresponding to each scale; Indicates the first The key vector corresponding to each scale; Indicates the first Each scale corresponds to a query mapping weight; Indicates the first The key mapping weights corresponding to each scale; Indicates the first Each scale corresponds to a query mapping bias term; Indicates the first Each scale corresponds to a key mapping bias term; Inter-scale similarity calculation: The first scale... The query vector corresponding to each scale With the The key vector corresponding to each scale Perform a dot product operation and normalize using a dimension factor to obtain the ... The first scale and the first Similarity scores between each scale : ； in, Indicates the first The first scale and the first Similarity scores between each scale; This represents a unified dimension for both the query vector and the key vector. Used for scaling the dot product result; Normalization processing: for the first The similarity scores between each scale are then Softmax normalized to obtain the i-th scale. The scale relation weights corresponding to the i-th scale; for the i-th scale... The first scale and the first The scale, the scale relation matrix A of the nth scale. Line 1 The elements of a column are represented as follows: ； in, Indicates the first The scale-coded features are used for calibration. Reference weights for each scale-coded feature; Indicates the first The first scale and the first Similarity scores between each scale; Indicates the first The first scale and the first Similarity scores between each scale; Through the above normalization process, the first The sum of the four association weights corresponding to each scale is 1, so that the similarity scores between different scales are transformed into a weight distribution for feature calibration; by all Composition of scale relation matrix A: ； in, This represents a query matrix composed of query vectors corresponding to four scales; This represents a key matrix composed of key vectors corresponding to four scales; This represents the unified dimension of the query vector and the key vector.

6. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 1, characterized in that, In step S4, the four scale calibration features obtained in step S3 are input into the decoder, and the decoder outputs the final decoded features, including: Third-scale decoding: Upsample the fourth-scale calibration features so that the spatial dimensions of the upsampled fourth-scale calibration features are consistent with those of the third-scale calibration features; concatenate the upsampled features and the third-scale calibration features along the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the third-scale fused features; guide the refinement and fusion of the third-scale fused features to obtain the third-scale decoded features. Second-scale decoding: Upsample the third-scale decoding features so that the spatial dimensions of the upsampled third-scale decoding features are consistent with the second-scale calibration features; concatenate the upsampled features and the second-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the second-scale fusion features; guide the refinement fusion of the second-scale fusion features to obtain the second-scale decoding features. First-scale decoding: Upsample the second-scale decoding features so that the spatial dimensions of the upsampled second-scale decoding features are consistent with the first-scale calibration features; concatenate the upsampled features and the first-scale calibration features in the channel dimension, and perform 1×1 convolution, batch normalization and ReLU activation to obtain the first-scale fusion features; guide the refinement fusion of the first-scale fusion features to obtain the first-scale decoding features, and use the first-scale decoding features as the final decoding features.

7. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 6, characterized in that, The specific process of rapid guided refinement fusion is as follows: The current scale fusion features are used as guiding features, and the features before upsampling in this decoding stage are used as features to be refined. The linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined in the local window. The current scale fusion features are then linearly reconstructed using the linear transformation coefficients and bias terms to obtain the decoding features of the current decoding stage.

8. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 7, characterized in that, The linear transformation coefficients and bias terms are calculated based on the response change relationship between the guiding features and the features to be refined within the local window. The linear transformation coefficients and bias terms are then used to perform linear reconstruction on the current scale fusion features to obtain the decoding features for the current decoding stage, including: Using bilinear interpolation downsampling, the height and width of the guiding feature are adjusted to match the feature to be refined, resulting in a size-adjusted guiding feature. The response change relationship within a local window is statistically analyzed between the guided feature after size adjustment and the feature to be refined, and the linear transformation coefficients and bias terms are calculated based on the response change relationship; the linear transformation coefficients and bias terms are represented in the form of a feature map, and the height and width of the feature map are consistent with the feature to be refined; Bilinear interpolation upsampling is performed on the feature maps representing linear transform coefficients and bias terms so that the height and width of the upsampled feature maps of linear transform coefficients and bias terms are consistent with the guiding features; Linear reconstruction of the current scale-fused features is performed using the upsampled linear transform coefficients and bias term to obtain the decoding features for the current decoding stage: ； in, Indicates the guiding features at the current scale; This indicates that the height and width have been adjusted to match. Consistent linear transformation coefficients; This indicates that the height and width have been adjusted to match. Consistent bias terms; This indicates the decoding characteristics at the current decoding stage; This indicates element-wise multiplication.

9. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 1, characterized in that, The specific process of the Sigmoid mapping is as follows: A nonlinear mapping is applied to the single-channel predicted features obtained by the 1×1 convolution mapping to obtain a retinal vessel probability map; the value of each pixel in the retinal vessel probability map ranges from 0 to 1, representing the probability that the pixel belongs to a vascular region; the Sigmoid mapping is expressed as: ； in, This indicates the single-channel predicted feature at the pixel location. The response value at that location, Indicates pixel position The probability that it belongs to a vascular region.

10. The retinal vessel image segmentation method based on orientation selection and scale relationship calibration according to claim 1, characterized in that, The specific process of threshold binarization is as follows: The retinal vessel probability map is judged based on a preset threshold to obtain a binary segmentation map of the retinal vessels; the threshold binarization process is expressed as follows: ； in, Indicates pixel position The binary segmentation result at the location, Indicates pixel position The corresponding pixel belongs to the blood vessel region. Indicates pixel position The corresponding pixels belong to the background area; This represents a preset threshold; when the probability of a blood vessel corresponding to a pixel location is greater than or equal to the preset threshold, the pixel is identified as a blood vessel region; when the probability of a blood vessel corresponding to a pixel location is less than the preset threshold, the pixel location is... The corresponding pixels are identified as the background area.

Images