Fundus blood vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating

Abstract

The invention discloses a fundus blood vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating. The fundus blood vessel segmentation method comprises the following steps: step 1, extracting multi-level features from an input image by using an encoder based on the deformable polar coordinate convolution; 2, enhancing the characteristics of the encoder by adopting a multi-scale gating attention module; and step 3, generating a blood vessel segmentation map by using the feature pyramid module based on the deformable polar coordinate convolution. According to the method, a deformable polar coordinate convolution is provided, and the offset of the deformable convolution is learned in a polar coordinate system, so that the deformable convolution can better adapt to the complex form of the blood vessel. Besides, in order to cope with the adjustment of large blood vessel scale difference and high similarity between blood vessels and background noise, the invention provides a multi-scale gating attention module, through a multi-scale strategy and a gating attention mechanism, the characterization capability of the model to multi-scale information and the anti-interference capability of the model to background noise are effectively enhanced, and the accuracy of the model is improved. Therefore, high-precision fundus blood vessel segmentation is realized.

Description

Technical Field

[0001] The present invention relates to a fundus blood vessel segmentation method, and in particular to a fundus blood vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating. Background Art

[0002] With the continuous advancement of medical imaging technology, color fundus images have gradually become an important tool for ophthalmologists to monitor eye health and diagnose fundus diseases. They can intuitively display the distribution of fundus blood vessels and the location of lesions, providing doctors with key information for diagnosis and treatment. Common fundus diseases, such as diabetic retinopathy and glaucoma, affect the structure and distribution of fundus blood vessels. Morphological changes in blood vessels, such as dilation, torsion, stenosis, or occlusion, are often direct manifestations of various fundus diseases. For example, in diabetic retinopathy, the formation of microaneurysms and new blood vessels is a sign of early lesions of the disease. By accurately segmenting fundus blood vessels, doctors can detect these subtle changes, allowing them to take early intervention measures to prevent further deterioration of the disease.

[0003] Fundus vessels exhibit a winding, bifurcated tubular structure, making it difficult for traditional convolutional models to adapt to their geometric diversity. Although deformable convolutions enhance flexibility by dynamically adjusting sampling positions, their offset learning lacks explicit constraints on vascular morphology and is susceptible to noise.

[0004] Vessel scales in fundus images vary significantly, with the distal ends of large vessels often breaking down into micron-sized branches. Furthermore, noise from pathological exudation and bleeding further interferes with the model's ability to identify vessels. While traditional attention mechanisms can suppress background noise, their single-scale feature fusion tends to overlook the multi-scale characteristics of blood vessels. Summary of the Invention

[0005] In order to alleviate the problem that the rigid receptive field of traditional convolutional neural networks is difficult to adapt to the geometric diversity of blood vessels and that the offset of deformable convolution is easily interfered by noise in an unconstrained Cartesian coordinate system, the present invention provides a fundus vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating. The method proposes a deformable polar coordinate convolution, which learns the offset of the deformable convolution in a polar coordinate system so that the deformable convolution can better adapt to the complex morphology of blood vessels. In addition, in order to cope with the adjustment of large differences in blood vessel scales and high similarity between blood vessels and background noise, the present invention proposes a multi-scale gated attention module. Through the multi-scale strategy and gated attention mechanism, the model's ability to represent multi-scale information and the model's ability to resist background noise are effectively enhanced, thereby achieving high-precision fundus vessel segmentation.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A method for segmenting fundus blood vessels based on deformable polar coordinate convolution and multi-scale gating includes the following steps:

[0008] Step 1: Use an encoder based on deformable polar convolution to extract multi-level features from the input image, where:

[0009] The encoder based on deformable polar coordinate convolution is implemented by replacing the ordinary convolution in the U-Net encoder with a deformable polar coordinate convolution module. The deformable polar coordinate convolution module includes ordinary convolution and deformable polar coordinate convolution. The input features are respectively subjected to ordinary convolution and deformable polar coordinate convolution, and the output of the deformable polar coordinate convolution module is obtained by adding them.

[0010] First, K center =(x center ,y center ) is converted from the Cartesian coordinate system to the polar coordinate system, and the center position coordinates in the polar coordinate system (r center ,θ center ), the specific conversion formula is as follows:

[0011]

[0012] where K center =(x center ,y center ) is the center coordinate of the convolution coordinate K in the Cartesian coordinate system, ρ center =(r center ,θ center ) is the coordinate of its center position in the polar coordinate system;

[0013] Introducing a deformation offset Δ in the convolution, the deformation offset Δ is split into two independent components: radial offset r offset and angular offset θ offset , for radial offset r offset and the angular offset θ offset Perform morphological constraints to ensure r offset The value range is between -1 and 1, θ offset The value range is between -π and π; after introducing the learnable deformation offset Δ, the position coordinates of the sliding window are expressed as:

[0014]

[0015] Among them, r offset ∈[-1,1] is the radial offset, θ offset ∈[-π,π] is the angle offset, ρ new =(r new ,θ new ) is the position coordinate of the sliding window, ρ center =(rcenter ,θ center ) is the coordinate of the center of the convolution in the polar coordinate system;

[0016] Convert the sampling point position coordinates back to the Cartesian coordinate system. The conversion formula is as follows:

[0017]

[0018] Among them, K new Represents the sampling point position coordinates in the Cartesian coordinate system;

[0019] Finally, the operational formula of deformable polar coordinate convolution is as follows:

[0020]

[0021] Among them, y(K center ) is K center The output of the deformable polar convolution in coordinates, k new Indicates the sampling point position coordinates after offset, x(k new ) represents the pixel value corresponding to the sampling point after the offset, w(k new ) The convolution kernel value corresponding to the sampling point;

[0022] Step 2: Use a multi-scale gated attention module to enhance encoder features, where:

[0023] The multi-scale gated attention module includes two parts: a multi-scale strategy and a gated attention module;

[0024] For the multi-scale strategy, the multi-scale gated attention module fuses the feature maps of each level in the encoder by splicing to obtain the multi-scale feature MS. Then, the generated MS is convolved to adjust the number of channels to make it the same as the encoder feature of the level, and then added and fused to obtain an input of the attention gating module;

[0025] For the gated attention module, the gated attention module receives two inputs: the first input is the encoder feature x enhanced by the multi-scale strategy, called the skip signal; the second input is the deep feature g, called the gate signal; the gate signal g is upsampled to match the size of the skip signal x; after the size is adjusted, the spatial information contained in the skip signal x and the deep feature information contained in the gate signal g are integrated through ordinary convolution and simple addition operations; then, the combined feature map is passed through a series of activation functions and filters to generate the attention weight w; finally, the attention weight w is multiplied by the skip signal x to achieve the weighting of the skip signal x. The specific formal formula is as follows:

[0026]

[0027] in, φ skip and φ gate is ordinary convolution, Relu is the activation function, and x′ is the output;

[0028] Step 3: Generate a blood vessel segmentation map using a feature pyramid module based on deformable polar coordinate convolution, where:

[0029] The feature pyramid module based on deformable polar coordinate convolution includes two parts: a deformable polar coordinate convolution module and a feature pyramid module;

[0030] The feature pyramid module uses a deformable polar coordinate convolution module to perform top-down fusion of multi-level features. The fused feature map is upsampled to the same size as the final output segmentation map. The feature maps after each layer are cascaded to form a new feature map, which is convolved to obtain the final segmentation result. In order to promote the fusion of multi-level features, the feature maps after each layer are amplified through 1×1 convolution to generate independent segmentation results for each layer, and the segmentation results of each layer are supervised.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] Inspired by the biological properties of blood vessels, this paper proposes a deformable polar coordinate convolution module, which models vascular structures as radial and angular offsets in a polar coordinate system. By constraining the offset range, this module can more naturally fit the vascular morphology, enhancing the ability to capture local features while suppressing invalid deformations, thereby improving vessel segmentation accuracy.

[0033] 2. The present invention designs a multi-scale gated attention module, which fuses the multi-scale features of the encoder at the jump connection, guides the model to focus on the vascular area through the gating signal, and achieves highly robust segmentation results in complex pathological scenes.

[0034] 3. The present invention learns the offset of the deformable convolution in a polar coordinate system so that the deformable convolution can better adapt to the complex morphology of blood vessels.

[0035] 4. This invention effectively enhances the model's ability to represent multi-scale information by dynamically fusing multi-level features into encoder features. Furthermore, it innovatively introduces an attention mechanism to adaptively filter the fused features. This recalibrates the weights of the feature space dimensions, effectively suppressing interference from redundant information and background noise while retaining key discriminative information. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1Flowchart of the fundus vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating.

[0037] Figure 2 Schematic diagram of deformable polar coordinate convolution.

[0038] Figure 3 Schematic diagram of the multi-scale gated attention module. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0040] The present invention provides a fundus vessel segmentation method based on deformable polar coordinate convolution and multi-scale gating. The method first uses an encoder based on deformable polar coordinate convolution to extract multi-level features from the input image, then uses an attention module containing multi-scale gating to enhance the features of the encoder, and finally uses a feature pyramid module based on deformable polar coordinate convolution to generate the final vessel segmentation prediction map. Figure 1 As shown, the specific steps include:

[0041] Step 1: Use an encoder based on deformable polar convolution to extract multi-level features from the input image.

[0042] In this step, by replacing the ordinary convolution in the U-Net encoder with a deformable polar coordinate convolution module, the radial and angular offsets are constrained in the polar coordinate system, and the sampling position is dynamically adjusted to adapt to the morphology of the blood vessels, thereby enhancing the representation ability of local geometric features.

[0043] In this step, the implementation of the encoder based on deformable polar coordinate convolution is to replace the ordinary convolution operation in the U-Net encoder with the deformable polar coordinate convolution module, such as Figure 1 As shown in the figure, since the U-Net encoder is widely used, this article focuses on the deformable polar convolution module, which includes both normal convolution and deformable polar convolution. The output of this module is obtained by applying normal convolution and deformable polar convolution to the input and adding them together.

[0044] The present invention regards the vascular structure as a series of overlapping elliptical or circular curves. This structure can be represented more concisely and intuitively in a polar coordinate system, such as Figure 2Compared to the traditional Cartesian coordinate system, the polar coordinate system has a natural advantage in representing geometric figures with rotational or central symmetry. For example, features such as vascular curvature and bifurcation can be simplified to radial and angular changes in the polar coordinate system, making it an ideal choice for analyzing and processing vascular images.

[0045] In order to realize the above idea, we first convert it from the Cartesian coordinate system to the polar coordinate system to obtain the center position coordinates (r center ,θ center ), the specific conversion formula is as follows:

[0046]

[0047] where K center =(x center ,y center ) is the center coordinate of the convolution coordinate K in the Cartesian coordinate system, ρ center =(r center ,θ center ) are the coordinates of its center position in the polar coordinate system.

[0048] At this time, the local area of ​​the fundus blood vessels can be regarded as a series of radial distributions around the center point, which not only reduces the computational complexity but also is consistent with the biological structural characteristics of the blood vessels.

[0049] Deformable convolution introduces a learnable offset, giving the convolution kernel the ability to dynamically adjust the sampling position, thereby more effectively capturing image features. To give the convolution kernel more flexibility and enable it to focus on the complex geometric features of vascular structures, this paper, inspired by deformable convolution, introduces a deformation offset Δ.

[0050] like Figure 2 As shown in the figure, the dotted dots represent the fixed sampling positions of the traditional convolution, and the solid dots are the sampling positions of the deformable polar coordinate convolution. Since the current system is in the polar coordinate system, the present invention splits the deformation offset Δ into two independent components: radial offset r offset and the angular offset θ offset This split allows the present invention to control the deformation of the convolution kernel in the radial and angular directions respectively, so as to better adapt to the characteristics of blood vessels such as curvature and bifurcation. However, if the model is allowed to learn deformation offset completely freely, it may cause it to deviate from the segmentation target and learn invalid features, thus affecting the model performance. To avoid this potential problem, the present invention adopts a constraint strategy, that is, the radial offset r offset and the angular offset θ offset Perform certain morphological constraints.

[0051] Specifically, the present invention ensures that offsetThe value range is between -1 and 1, θ offset The value range is between -π and π. This constraint not only prevents excessive deformation and ensures that the perception range does not spread disorderly, but also ensures the stability and continuity of the offset. In this way, the deformation offset Δ can better fit the natural geometric characteristics of the vascular structure, thereby improving the segmentation accuracy of the model. After introducing the learnable deformation offset Δ, the position coordinates of the sliding window can be expressed as:

[0052]

[0053] Among them, r offset ∈[-1,1] is the radial offset, θ offset ∈[-π,π] is the angle offset, ρ new =(r new ,θ new ) is the position coordinate of the sliding window, ρ center =(r center ,θ center ) is the center coordinate of the convolution in the polar coordinate system.

[0054] Through the above steps, the present invention obtains a series of sampling point position coordinates that are more consistent with the fundus vascular structure. These sampling points are optimized in polar coordinates and can more accurately capture the geometric characteristics of the blood vessels than fixed sampling points. In order to apply these sampling points to the standard convolution operation, the present invention converts them back to the Cartesian coordinate system. The conversion formula is as follows:

[0055]

[0056] Among them, K new Represents the sampling point position coordinates in the Cartesian coordinate system.

[0057] Finally, the operational formula of deformable polar coordinate convolution is as follows:

[0058]

[0059] Among them, y(K center ) is K center The output of the deformable polar convolution in coordinates, k new Indicates the sampling point position coordinates after offset, x(k new ) represents the pixel value corresponding to the sampling point after the offset, w(k new )The convolution kernel value corresponding to the sampling point.

[0060] Step 2: Use a multi-scale gated attention module to enhance encoder features.

[0061] In this step, the multi-level features of the splicing encoder are used to generate aggregated multi-scale features. Combined with the attention mechanism guided by the gating signal, the features of the jump connection are adaptively weighted in the spatial dimension, effectively suppressing background noise and highlighting the discriminative information of the vascular area.

[0062] like Figure 3 As shown in Figure 2, the multi-scale gated attention module consists of two main components: a multi-scale strategy and a gated attention module. For the multi-scale strategy, this module concatenates and fuses the feature maps at each encoder level to generate multi-scale features (MSs). The generated MSs are then convolved to adjust the number of channels to match the encoder features at that level. These MSs are then added together to form an input to the attention gating module.

[0063] For the gated attention module, the module receives two inputs: the first input is the encoder feature x enhanced by the multi-scale strategy, called the skip signal; the second input is the deep feature g, called the gating signal. Since the dimensions of the gating signal g are inconsistent with the skip signal x, the present invention first upsamples the gating signal g to match the size of the skip signal x. After the size is adjusted, the spatial information contained in the skip signal x and the deep feature information contained in the gating signal g are integrated through ordinary convolution and simple addition operations. Subsequently, the present invention passes the combined feature map through a series of activation functions and filters to generate an attention weight w. Finally, the weight w is multiplied by the skip signal x to achieve weighting of the skip signal x. In this way, the model can effectively highlight the target area containing blood vessels while suppressing other irrelevant areas. The specific formal formula is as follows:

[0064]

[0065] in, φ skip and φ gate is a normal convolution, Relu is the activation function, and x' is the output.

[0066] Step 3: Generate a blood vessel segmentation map using a feature pyramid module based on deformable polar coordinate convolution.

[0067] In this step, the deformable polar coordinate convolution module is used to extract features, and multi-scale feature fusion is performed in a top-down manner. Then, the feature maps at different levels are upsampled and cascaded, and finally the final segmentation result is generated through convolution.

[0068] The feature pyramid module based on deformable polar coordinate convolution mainly includes two parts, namely the deformable polar coordinate convolution module and the feature pyramid module. For the feature pyramid module, the present invention adopts a structure similar to the feature pyramid network (FPN), such as Figure 1 As shown in the figure, the multi-level features are first fused top-down using a deformable polar convolution module. The fused feature map is then upsampled to the same size as the final segmentation map. Subsequently, the invention concatenates the amplified feature maps of each layer to form a new feature map. This new feature map contains rich information from different scales. Finally, this feature map is convolved to obtain the final segmentation result.

[0069] In addition, in order to promote the fusion of multi-level features, the present invention refers to the design of FPN and further sends the amplified feature map of each layer into the convolution module for processing through 1×1 convolution, generating independent segmentation results for each layer and supervising the segmentation results of each layer.

[0070] Step 4: Design a composite loss function to train the model end-to-end.

[0071] In this step, the segmentation accuracy and continuity of the model are optimized by combining binary cross entropy loss (BCE), Dice loss, and topology-preserving clDice loss. At the same time, the intermediate layer output of the feature pyramid module is supervised to promote multi-scale feature fusion.

[0072] In the training of the model, the goal of the present invention is to design a loss function that can significantly improve the segmentation accuracy and segmentation continuity of the model. In terms of segmentation accuracy, the evaluation of blood vessel segmentation accuracy not only depends on the classification accuracy of blood vessels and background, but also needs to consider the degree of overlap between the predicted value and the true value. Therefore, the present invention combines the binary cross entropy loss (BCE) function and the Dice loss function to ensure the performance of the model in multiple key aspects. In terms of segmentation continuity, the present invention uses the clDice loss function. This loss function is a new type of topology-preserving loss function designed for the segmentation of tubular structures (such as blood vessels, neurons, and roads). It aims to maintain the integrity of the topological structure and optimize the segmentation results by calculating the Dice coefficient of the intersection of the centerline and the volume. The basic loss function Loss used in the present invention base for:

[0073] Loss base (Y,G)=BCE(Y,G)+0.5Dice(Y,G)+0.5clDice(Y,G)

[0074] Among them, Y is the predicted value and G is the true value.

[0075] In order to promote the fusion of multi-level features, in the feature pyramid module based on polar coordinate convolution, the present invention supervises the aggregated features of each level, such as Figure 1 As shown, the composite loss function Loss is:

[0076] Loss = 0.8 Loss base (Y1,G)+0.1Loss base (Y2,G)+0.1Loss base (Y3,G)

[0077] Among them, Y1 is the final predicted blood vessel segmentation map, and Y1 and Y2 are blood vessel segmentation maps predicted by intermediate layer features.

[0078] The performance of the present invention is evaluated on two popular datasets, namely DRIVE and CHASEDB1. The present invention uses three evaluation criteria, namely accuracy (ACC), mean intersection over union (mIoU) and Dice Similarity Coefficient (DSC). Accuracy indicates the proportion of all pixels in the image that are correctly classified, that is, the proportion of pixels correctly classified as vascular pixels and background pixels to all pixels, which can represent the overall segmentation performance of the model. The Dice coefficient is used to measure the similarity between two sample sets. The value range is between [0,1]. The closer the value is to 1, the higher the degree of consistency between the predicted result and the true label. In color fundus images, since the proportion of vascular pixels is much smaller than that of background pixels, class imbalance problems will occur during segmentation. In this case, the Dice coefficient will pay more attention to the overlap between the predicted results and the true label than the accuracy, rather than simply focusing on the proportion of the category. mIoU measures model performance by calculating the ratio of the intersection over union of the predicted results and the true labels. It can comprehensively consider the pixel-level prediction accuracy of the model on all categories and fairly evaluate the model when dealing with objects of different categories and sizes.

[0079] Table 1 shows a performance comparison between the present invention and the state-of-the-art methods on the DRIVE dataset. Table 2 shows a performance comparison between the present invention and the state-of-the-art methods on the CHASEDB1 dataset. As can be seen from Tables 1 and 2, the present invention achieves state-of-the-art results on all evaluation metrics on both datasets, demonstrating that the present invention not only addresses existing challenges in fundus vessel segmentation but also outperforms existing methods.

[0080] Table 1 Quantitative comparison on the DRIVE dataset. Unit (%)

[0081] method ACC mIoU DSC U-Net 95.2 81.3 80.5 U-Net++ 95.8 81.7 81.2 Deformable-ConvNet 96.2 81.2 78.8 DUNet 96.5 81.7 81.7 DSCNet 96.9 82.0 81.8 AttentionU-Net 96.1 81.4 80.9 CAMU-Net 96.7 82.0 82.4 The present invention 97.3 82.7 82.4

[0082] Table 2 Quantitative comparison on the CHASEDB1 dataset. Unit (%)

[0083] method ACC mIoU DSC U-Net 96.3 81.2 80.3 U-Net++ 96.4 81.0 78.9 Deformable-ConvNet 96.5 80.5 73.5 DUNet 96.8 80.8 78.5 DSCNet 97.1 81.8 80.6 AttentionU-Net 96.7 81.3 81.0 CAMU-Net 97.0 81.6 80.9 The present invention 97.7 82.5 81.5

Claims

1. A method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating, characterized by The method comprises the following steps: Step 1: Use an encoder based on deformable polar convolution to extract multi-level features from the input image; Step 2: Use a multi-scale gated attention module to enhance encoder features; Step 3: Generate a blood vessel segmentation map using a feature pyramid module based on deformable polar coordinate convolution.

2. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 1 is characterized in that The encoder based on deformable polar coordinate convolution is implemented by replacing the ordinary convolution in the U-Net encoder with a deformable polar coordinate convolution module. The deformable polar coordinate convolution module includes ordinary convolution and deformable polar coordinate convolution. The input features are respectively subjected to ordinary convolution and deformable polar coordinate convolution, and the output of the deformable polar coordinate convolution module is obtained by adding them.

3. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 1 is characterized in that The calculation formula of the deformable polar coordinate convolution is as follows: First, K center =(x center ,y center ) is converted from the Cartesian coordinate system to the polar coordinate system, and the center position coordinates in the polar coordinate system (r center ,θ center ), the specific conversion formula is as follows: where K center =(x center ,y center ) is the center coordinate of the convolution coordinate K in the Cartesian coordinate system, ρ center =(r center ,θ center ) is the coordinate of its center position in the polar coordinate system; Introducing a deformation offset Δ in the convolution, the deformation offset Δ is split into two independent components: radial offset r offset and angular offset θ offset , for radial offset r offset and angular offset θ offset Perform morphological constraints to ensure r offset The value range is between -1 and 1, θ offset The value range is between -π and π; after introducing the learnable deformation offset Δ, the position coordinates of the sliding window are expressed as: Among them, r offset ∈[-1,1] is the radial offset, θ offset ∈[-π,π] is the angle offset, ρ new =(r new ,θ new ) is the position coordinate of the sliding window, ρ center =(r center ,θ center ) is the coordinate of the center of the convolution in the polar coordinate system; Convert the sampling point position coordinates back to the Cartesian coordinate system. The conversion formula is as follows: Among them, K new Represents the sampling point position coordinates in the Cartesian coordinate system; Finally, the operational formula of deformable polar coordinate convolution is as follows: Among them, y(K center ) is K center The output of the deformable polar convolution in coordinates, k new Indicates the sampling point position coordinates after offset, x(k new ) represents the pixel value corresponding to the sampling point after the offset, w(k new )The convolution kernel value corresponding to the sampling point.

4. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 1 is characterized in that The multi-scale gated attention module includes two parts: a multi-scale strategy and a gated attention module; For the multi-scale strategy, the multi-scale gated attention module fuses the feature maps of each level in the encoder by splicing to obtain the multi-scale feature MS. Then, the generated MS is convolved to adjust the number of channels to make it the same as the encoder feature of the level, and then added and fused to obtain an input of the attention gating module; For the gated attention module, the gated attention module receives two inputs: the first input is the encoder feature x enhanced by the multi-scale strategy, called the skip signal; the second input is the deep feature g, called the gate signal; the gate signal g is upsampled to match the size of the skip signal x; After resizing, the spatial information contained in the skip signal x and the deep feature information contained in the gated signal g are integrated through ordinary convolution and simple addition operations; Subsequently, the combined feature map is passed through a series of activation functions and filters to generate the attention weight w. Finally, the attention weight w is multiplied by the jump signal x to achieve the weighting of the jump signal x. The specific formal formula is as follows: in, φ skip and φ gate is a normal convolution, Relu is the activation function, and x' is the output.

5. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 1 is characterized in that The feature pyramid module based on deformable polar coordinate convolution includes two parts: a deformable polar coordinate convolution module and a feature pyramid module. The feature pyramid module uses the deformable polar coordinate convolution module to perform top-down fusion of multi-level features, and the fused feature map is enlarged to the same size as the final output segmentation map through upsampling operation. The feature map after each layer is enlarged is cascaded to form a new feature map, and the feature map is convolved to obtain the final segmentation result. In order to promote the fusion of multi-level features, the feature map after each layer is enlarged is subjected to 1×1 convolution to generate an independent segmentation result for each layer, and the segmentation result of each layer is supervised.

6. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 1, characterized in that The method designs a composite loss function to perform end-to-end training on the model. The composite loss function Loss is: Loss=0.8Loss base (Y1,G)+0.1Loss base (Y2,G)+0.1Loss base (Y3,G) Among them, Loss base is the basic loss function, Y1 is the final predicted blood vessel segmentation map, and Y1 and Y2 are the blood vessel segmentation maps predicted by the intermediate layer features.

7. The method for fundus vessel segmentation based on deformable polar coordinate convolution and multi-scale gating according to claim 6, characterized in that The basic loss function Loss base for: Loss base (Y,G)=BCE(Y,G)+0.5Dice(Y,G)+0.5clDice(Y,G) Among them, Y is the predicted value and G is the true value.

Images