Segmentation and extraction method of retinal blood vessel
By introducing multi-core channel spatial attention module, pixel-level self-attention network and dynamic adaptive convolution module in retinal vascular segmentation, the problem of differences in traditional network processing of feature information at different scales is solved, the segmentation effect is improved, and complexity is achieved Accurate capture and segmentation of retinal blood vessels.
Patent Information
- Application Number
- CN202410935284.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-15
AI Technical Summary
The existing retinal vascular segmentation algorithm has characterization differences when processing characteristic information at different scales. The traditional U-type encoder-decoder network cannot effectively process shallow high resolution and deep high-dimensional semantic information, resulting in poor segmentation effect.
A multi-core channel spatial attention module, pixel-level self-attention network module and dynamic adaptive convolution module are used to build a U-shaped symmetric network, and the combined optimization of cross entropy loss and dice loss can improve the vascular feature learning ability and segmentation accuracy.
It improves the ability to learn vascular characteristics, enhances the network's ability to segment complex retinal blood vessels, captures variable vascular structures, and significantly improves the segmentation effect.
Smart Images

Figure CN118967708A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of ophthalmic imaging, and in particular to a method for segmenting and extracting retinal blood vessels. Background Art
[0002] Retinal vessel segmentation is an important non-destructive medical imaging task in computer vision and is crucial for diagnosing fundus diseases. Although deep learning-based methods dominate this field, existing segmentation algorithms mainly utilize U-shaped encoder-decoder networks with skip connections. However, there are obvious representation differences between feature information at different scales. Specifically, the feature maps at the shallow stage have high resolution and low dimension, which can provide the network with fine-grained vascular information. With multiple downsampling, the resolution gradually decreases. At this time, the feature maps at the deep stage contain rich high-dimensional semantic information, and the traditional U-shaped structure cannot discretely process these two types of features. Summary of the invention
[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for segmenting and extracting retinal blood vessels.
[0004] The object of the present invention is achieved by the following technical solution: A method for segmenting and extracting retinal blood vessels comprises the following steps:
[0005] S1: Acquire retinal fundus vascular data and preprocess the data;
[0006] S2: Construct multi-core channel spatial attention module, pixel-level self-attention network module and dynamic adaptive convolution module;
[0007] S3: forming a U-shaped symmetrical network according to the modules in step S2;
[0008] S4: Use the weighted combination of cross entropy loss and dice loss as the objective function of the network, perform multiple iterations of optimization, and obtain the optimal weight of the network;
[0009] S5: Obtain blood vessel segmentation results based on the model.
[0010] Preferably, in step S1, the data is cropped, flipped, translated, rotated and normalized, and the input image is scaled to a resolution of 256×256.
[0011] Preferably, the multi-core channel spatial attention module includes a multi-core feature extraction submodule, a channel attention submodule and a spatial attention submodule;
[0012] The multi-core feature extraction submodule generates multiple feature maps, denoted as x1, x2, and x3, through parallel input of convolution kernels of different scales;
[0013] The channel attention submodule is used to aggregate feature maps, and then average pooling, full connection and activation function are performed in sequence to obtain attention weights;
[0014] The spatial attention submodule is used to enhance the spatial information of features.
[0015] Preferably, the processing steps of the channel attention submodule are as follows:
[0016] A1: x2 and x3 are sent to the channel refinement path in parallel for aggregation, and then average pooling, full connection and activation function are performed in sequence to obtain attention weights w1 and w2;
[0017] A2: Attention weights w1 and w2 are multiplied by x2 and x3 respectively to obtain new feature maps x'2 and x'3. The calculation formula is:
[0018]
[0019]
[0020] Among them, FC is the fully connected layer;
[0021] A3: Perform aggregation operations on x'2 and x'3 to determine the importance of each channel.
[0022]
[0023] Among them, Map channel It is the feature aggregation graph after multi-core channel enhancement;
[0024] The processing steps of the spatial attention submodule are as follows:
[0025] B1: Use deep convolution with a kernel size of 7×7 to extract the spatial information of blood vessels.
[0026]
[0027]
[0028] Among them, w3 and w4 are a set of weights, w3 and w4 respectively represent the blood vessel position information of different feature maps in the feature space, and mean is the average operation in a given dimension;
[0029] B2: x1 is obtained through the depth separation convolution operation, and then input into the spatial refinement path in parallel with the channel enhanced feature map, and finally added to the original feature x.
[0030]
[0031]
[0032]
[0033]
[0034] Preferably, the steps of constructing a pixel-level self-attention network module are as follows:
[0035] C1: Convert the depth feature of size [256, 32, 32] into a pixel matrix of the same shape;
[0036] C2: The input features are converted into a pixel matrix of [b, 256, 32, 32] after pixel sampling, and then the pixel matrix is sent to the pixel embedding layer to obtain multiple pixel-level patches p. Then, learnable parameter position embedding is added to supplement the position information of the pixels in the label.
[0037] pixel_matrix=pixel_sampling(x);
[0038] p=pixel_embedding(pixel_matrix);
[0039] p' = Linear(x);
[0040] embedded patch =p+p';
[0041] Among them, x is the depth feature;
[0042] C3: The pixel-level patch p is fed into the Transformer encoding module to build a long-range global dependency network.
[0043] Y=MSA(embedded_patch)+embedded_patch;
[0044] Output = MLP (Y) + Y;
[0045] Among them, MSA is a multi-head self-attention module, MLP is multiple serial linear functions, and Output is the output of the pixel-level self-attention module.
[0046] Preferably, the size of the pixel-level patch p is [256, 1, 1].
[0047] Preferably, each coordinate point K±i, (i=2) in the grid coordinate G of the dynamic adaptive convolution module in the standard convolution process can be expressed as:
[0048] G=(-2,2),(-2,-1),...,(2,1),(2,2);
[0049] The predicted value corresponding to the input feature and position P0 can be expressed as:
[0050] y(p0)=∑(w(ki)+x(p0+Ki));
[0051] An offset δ is introduced based on the standard convolution. During the iterative generation of the offset, the position of the next offset generation depends on the position of the previous offset, and the offset range is [-1, 1].
[0052] y(p0)=∑(w(ki)+x(p0+Ki+δx))+∑(w(ki)+x(p0+Ki+δy));
[0053] Among them, δx is the offset in the x-axis direction, and δy is the offset in the y-axis direction.
[0054] Preferably, in step S3, the input of the multi-core channel spatial attention module is the features generated by the shallow encoder, and the features extracted by the multi-core channel spatial attention module are connected in series with the features generated by the decoder through jump connection; the pixel-level self-attention network module is embedded in the bottleneck of the encoder and the decoder, including a position embedding layer, a pixel sampling layer and four multi-head self-attention layers, the pixel sampling layer is used to convert the image features into pixel-level patches, and send them to the self-attention layer together with the position embedding layer to obtain pixel-level global dependencies; the dynamic adaptive convolution module is set between the last layer of the decoder and the segmentation head.
[0055] Preferably, in step S4, the network loss function for training is based on the dice coefficient and the cross entropy loss, which are respectively expressed as Dice and CE, and the CE loss is expressed as:
[0056] Lce=-Σlog(y true )*log(y pred );
[0057] Among them, y true is the actual label value, y pred is the probability value predicted by the network, log is the natural logarithm;
[0058] L dice The mathematical representation of is:
[0059]
[0060] Among them, Pred i is the predicted value, i is the index of each pixel, Ture i is the true label, smooth is the smoothing constant;
[0061] The total loss is:
[0062] Loss = λLce + Ldice;
[0063] Among them, λ is the loss weight, λ∈[0,1].
[0064] The present invention has the following advantages: the present invention improves the learning ability of vascular features through a multi-core channel spatial attention module and a pixel-level self-attention network module, and then enhances the network's ability to accurately segment complex retinal blood vessels through a dynamic adaptive convolution module, thereby capturing variable vascular structures. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 A schematic diagram of the structure of the present invention
[0066] Figure 2 It is a schematic diagram of the structure of the multi-core channel spatial attention module;
[0067] Figure 3 Schematic diagram of the structure of the pixel-level self-attention module;
[0068] Figure 4 It is a schematic diagram of a U-shaped symmetrical network structure;
[0069] Figure 5 Schematic diagram of the segmentation effects of different methods;
[0070] Figure 6 A schematic diagram showing the detailed comparison of segmentation results. DETAILED DESCRIPTION
[0071] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0072] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0073] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0074] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0075] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, or the positions or positional relationships in which the product of the invention is usually placed when in use, or the positions or positional relationships commonly understood by those skilled in the art, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.
[0076] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0077] In this embodiment, if Figure 1 As shown, a method for segmenting and extracting retinal blood vessels includes the following steps:
[0078] S1: Acquire retinal fundus vascular data and preprocess the data;
[0079] S2: Construct multi-core channel spatial attention module, pixel-level self-attention network module and dynamic adaptive convolution module;
[0080] S3: forming a U-shaped symmetrical network according to the modules in step S2;
[0081] S4: Use the weighted combination of cross entropy loss and dice loss as the objective function of the network, perform multiple iterations of optimization, and obtain the optimal weight of the network;
[0082] S5: Obtain the vascular segmentation results based on the model. The multi-core channel spatial attention module and the pixel-level self-attention network module are used to improve the learning ability of vascular features, and the dynamic adaptive convolution module is used to enhance the network's ability to accurately segment complex retinal vessels, thereby capturing variable vascular structures.
[0083] Furthermore, in step S1, the data is cropped, flipped, translated, rotated and normalized, and the input image is scaled to a resolution of 256×256 in order to align with the network size.
[0084] In this embodiment, if Figure 2 As shown, the multi-core channel spatial attention module includes a multi-core feature extraction submodule, a channel attention submodule and a spatial attention submodule;
[0085] The multi-core feature extraction submodule generates multiple feature maps, denoted as x1, x2, and x3, through parallel input of convolution kernels of different scales;
[0086] The channel attention submodule is used to aggregate feature maps, and then average pooling, full connection and activation function are performed in sequence to obtain attention weights;
[0087] The spatial attention submodule is used to enhance the spatial information of the feature. Furthermore, the processing steps of the channel attention submodule are as follows:
[0088] A1: x2 and x3 are sent to the channel refinement path in parallel for aggregation, and then average pooling, full connection and activation function are performed in sequence to obtain attention weights w1 and w2;
[0089] A2: Attention weights w1 and w2 are multiplied by x2 and x3 respectively to obtain new feature maps x'2 and x'3. The calculation formula is:
[0090]
[0091]
[0092] Among them, FC is the fully connected layer;
[0093] A3: Perform aggregation operations on x'2 and x'3 to determine the importance of each channel.
[0094]
[0095] Among them, Map channel It is the feature aggregation graph after multi-core channel enhancement;
[0096] The processing steps of the spatial attention submodule are as follows:
[0097] B1: Use deep convolution with a kernel size of 7×7 to extract the spatial information of blood vessels.
[0098]
[0099]
[0100] Among them, w3 and w4 are a set of weights, w3 and w4 respectively represent the blood vessel position information of different feature maps in the feature space, and mean is the average operation in a given dimension;
[0101] B2: x1 is obtained through the depth-wise separation convolution operation, and then input into the spatial refinement path in parallel with the channel-enhanced feature map, and finally added to the original feature x, in order to avoid the loss of original information.
[0102]
[0103]
[0104]
[0105]
[0106] Specifically, the spatial information of the feature reflects the position of the blood vessel in the receptive field. The larger the convolution kernel, the larger the range of the receptive field. In order to achieve a balance between computational complexity and performance, a deep convolution with a kernel size of 7×7 is used to extract the spatial information of the blood vessel. x1 is obtained through a deep separation convolution operation and then input into the spatial refinement path in parallel with the channel-enhanced feature map. In this way, the spatial information of the feature can be enhanced and the structure of the blood vessel can be effectively captured. The feature map of the multi-core channel spatial attention module is connected to the original feature through a residual structure, thereby strengthening the information supplement to the decoder and enhancing the representation ability of the network.
[0107] In this embodiment, if Figure 3 As shown in Figure 2, the steps to construct a pixel-level self-attention network module are as follows:
[0108] C1: Convert the depth feature of size [256, 32, 32] into a pixel matrix of the same shape;
[0109] C2: The input features are converted into a pixel matrix of [b, 256, 32, 32] after pixel sampling, and then the pixel matrix is sent to the pixel embedding layer to obtain multiple pixel-level patches p. Then, learnable parameter position embedding is added to supplement the position information of the pixels in the label.
[0110] pixel_matrix=pixel_sampling(x);
[0111] p=pixel_embedding(pixel_matrix);
[0112] p' = Linear(x);
[0113] embedded patch =p+p';
[0114] Where x is a deep feature, x belongs to [b, 256, 32, 32]. After feature flattening, its shape is [b, 256, 1024]. After feature dimension transformation, p is obtained, and its shape is [b, 256, 1024]. + is an addition operation, so that the shape of the embedded patch does not change, and its shape is still [1024, b, 256]. Preferably, the size of the pixel-level patch p is [256, 1, 1]. Specifically, this step treats each individual pixel as a patch to obtain the semantic correlation between each pixel, and then adds a learnable parameter position embedding to supplement the position information of the pixel in the tag. The position embedding p∈R 256×1×1 Available via a Linear layer.
[0115] C3: The pixel-level patch p is fed into the Transformer encoding module to build a long-range global dependency network.
[0116] Y=MSA(embedded_patch)+embedded_patch;
[0117] Output = MLP (Y) + Y;
[0118] Among them, MSA is a multi-head self-attention module, MLP is a number of serial linear functions, and Output is the output of the pixel-level self-attention module. Specifically, the number of self-attention heads is set to 4, and there are two linear layers in MLP.
[0119] In this embodiment, each coordinate point K±i, (i=2) in the grid coordinate G of the dynamic adaptive convolution module in the standard convolution process can be expressed as:
[0120] G=(-2,2),(-2,-1),...,(2,1),(2,2);
[0121] The predicted value corresponding to the input feature and position P0 can be expressed as:
[0122] y(p0)=Σ(w(ki)+x([0+Ki));
[0123] An offset δ is introduced based on the standard convolution. During the iterative generation of the offset, the position of the next offset generation depends on the position of the previous offset, and the offset range is [-1, 1].
[0124] y(p0)=Σ(w(ki)+x(p0+Ki+δx))+∑(w(ki)+x(p0+Ki+δy));
[0125] Among them, δx is the offset in the x-axis direction, and δy is the offset in the y-axis direction. Specifically, an additional offset is introduced in the standard convolution process, which is specifically designed to learn and adapt to curved vascular structures, so that it forms an adaptive receptive field, thereby flexibly capturing variable vascular structures, thereby enhancing the network's ability to accurately segment complex retinal vessels. In this embodiment, the offset δ is not generated randomly. The standard convolution kernel is first expanded along the x-axis and y-axis directions. In order to limit the offset drift, in the iterative generation process of the offset, the position of the next offset generation depends on the position of the previous offset, and the offset range is [-1,1]. Since the offset δ is not integer data, bilinear interpolation is used to generate data.
[0126] In this embodiment, if Figure 4 As shown, in step S3, the input of the multi-core channel spatial attention module is the feature generated by the shallow encoder. The features extracted by the multi-core channel spatial attention module are connected in series with the features generated by the decoder through jump connection, so as to obtain multi-scale fine-grained information; the pixel-level self-attention network module is embedded in the bottleneck of the encoder and decoder, including a position embedding layer, a pixel sampling layer and four multi-head self-attention layers. The pixel sampling layer is used to convert image features into pixel-level patches, and send them to the self-attention layer together with the position embedding layer to obtain pixel-level global dependencies; the dynamic adaptive convolution module is set between the last layer of the decoder and the segmentation head to adapt to the complex vascular topology.
[0127] In this embodiment, in step S4, the network loss function for training is based on the dice coefficient and the cross entropy loss, which are respectively expressed as Dice and CE. The CE loss is expressed as:
[0128] Lce=-∑log(y true )*log(y pred );
[0129] Among them, y true is the actual label value, y pred is the probability value predicted by the network, log is the natural logarithm;
[0130] L dice The mathematical representation of is:
[0131]
[0132] Among them, Pred i is the predicted value, i is the index of each pixel, Ture i is the true label, and smooth is a smoothing constant. Specifically, smooth is a very small constant, such as 1e-5, which prevents the denominator from being 0 and improves the stability of the calculation.
[0133] The total loss is:
[0134] Loss = λLce + Ldice;
[0135] Where λ is the loss weight, λ∈[0,1]. Preferably, λ is 0.5.
[0136] In this embodiment, the labeled retinal blood vessel samples and the labeled label images are input into the network together, and iterative optimization is performed through step S4. After the training is completed, the blood vessel segmentation structure is obtained. The results are input into the existing test sample prediction model and compared with other existing methods, such as Figure 5 and Figure 6 As shown, the network can accurately and effectively segment the patient's retinal blood vessels.
[0137] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for segmenting and extracting retinal blood vessels, characterized in that: The following steps are involved: S1: Acquire retinal fundus vascular data and preprocess the data; S2: Construct multi-core channel spatial attention module, pixel-level self-attention network module and dynamic adaptive convolution module; S3: forming a U-shaped symmetrical network according to the modules in step S2; S4: Use the weighted combination of cross entropy loss and dice loss as the objective function of the network, perform multiple iterations of optimization, and obtain the optimal weight of the network; S5: Obtain blood vessel segmentation results based on the model.
2. The method for segmenting and extracting retinal blood vessels according to claim 1, characterized in that: In step S1, the data is cropped, flipped, translated, rotated and normalized, and the input image is scaled to a resolution of 256×256.
3. The method for segmenting and extracting retinal blood vessels according to claim 2, characterized in that: The multi-core channel spatial attention module includes a multi-core feature extraction submodule, a channel attention submodule and a spatial attention submodule; The multi-core feature extraction submodule generates multiple feature maps, denoted as x1, x2 and x3, through parallel input of convolution kernels of different scales; The channel attention submodule is used to aggregate the feature maps, and then average pooling, full connection and activation function are performed in sequence to obtain the attention weight; The spatial attention submodule is used to enhance the spatial information of features.
4. The method for segmenting and extracting retinal blood vessels according to claim 3, characterized in that: The processing steps of the channel attention submodule are as follows: A1: x2 and x3 are sent to the channel refinement path in parallel for aggregation, and then average pooling, full connection and activation function are performed in sequence to obtain attention weights w1 and w2; A2: Attention weights w1 and w2 are multiplied by x2 and x3 respectively to obtain new feature maps x'2 and x'3. The calculation formula is: Among them, FC is the fully connected layer; A3: Perform aggregation operations on x'2 and x'3 to determine the importance of each channel. Among them, Map channel It is the feature aggregation graph after multi-core channel enhancement; The processing steps of the spatial attention submodule are as follows: B1: Use deep convolution with a kernel size of 7×7 to extract the spatial information of blood vessels. Among them, w3 and w4 are a set of weights, w3 and w4 respectively represent the blood vessel position information of different feature maps in the feature space, and mean is the average operation in a given dimension; B2: x1 is obtained through the depth separation convolution operation, and then input into the spatial refinement path in parallel with the channel enhanced feature map, and finally added to the original feature x.
5. The method for segmenting and extracting retinal blood vessels according to claim 4, characterized in that: The steps of constructing the pixel-level self-attention network module are as follows: C1: Convert the depth feature of size [256, 32, 32] into a pixel matrix of the same shape; C2: The input features are converted into a pixel matrix of [b, 256, 32, 32] after pixel sampling, and then the pixel matrix is sent to the pixel embedding layer to obtain multiple pixel-level patches p. Then, learnable parameter position embedding is added to supplement the position information of the pixels in the label. pixel_matrix=pixel_sampling(x); p=pixel_embedding(pixel_matrix); p′=Linear(x); embedded patch =p+p′; Among them, x is the depth feature; C3: The pixel-level patch p is fed into the Transformer encoding module to build a long-range global dependency network. Y=MSA(embedded_patch)+embedded_patch; Output = MLP (Y) + Y; Among them, MSA is a multi-head self-attention module, MLP is multiple serial linear functions, and Output is the output of the pixel-level self-attention module.
6. The method for segmenting and extracting retinal blood vessels according to claim 5, characterized in that: The size of the pixel-level patch p is [256, 1, 1].
7. The method for segmenting and extracting retinal blood vessels according to claim 6, characterized in that: Each coordinate point K±i, (i=2) in the grid coordinate G of the dynamic adaptive convolution module in the standard convolution process can be expressed as: G=(-2,2),(-2,-1),……,(2,1),(2,2); The predicted value corresponding to the input feature and position P0 can be expressed as: y(p0)=∑(w(ki)+x(p0+Ki)); An offset δ is introduced based on the standard convolution. During the iterative generation of the offset, the position of the next offset generation depends on the position of the previous offset, and the offset range is [-1, 1]. y(p0)=∑(w(ki)+x(p0+Ki+δx))+∑(w(ki)+x(p0+Ki+δy)); Among them, δx is the offset in the x-axis direction, and δy is the offset in the y-axis direction.
8. The method for segmenting and extracting retinal blood vessels according to claim 7, characterized in that: In step S3, the input of the multi-core channel spatial attention module is the features generated by the shallow encoder, and the features extracted by the multi-core channel spatial attention module are connected in series with the features generated by the decoder through jump connection; the pixel-level self-attention network module is embedded in the bottleneck of the encoder and the decoder, including a position embedding layer, a pixel sampling layer and four multi-head self-attention layers, and the pixel sampling layer is used to convert the image features into pixel-level patches, and send them to the self-attention layer together with the position embedding layer to obtain pixel-level global dependencies; the dynamic adaptive convolution module is arranged between the last layer of the decoder and the segmentation head.
9. The method for segmenting and extracting retinal blood vessels according to claim 8, characterized in that: In step S4, the network loss function for training is based on the dice coefficient and the cross entropy loss, which are respectively expressed as Dice and CE. The CE loss is expressed as: Lce=-∑log(y true )*log(y pred ); Among them, y true is the actual label value, y pred is the probability value predicted by the network, log is the natural logarithm; L dice The mathematical representation of is: Among them, Pred i is the predicted value, i is the index of each pixel, Ture i is the true label, smooth is the smoothing constant; The total loss is: Loss = λLce + Ldice; Among them, λ is the loss weight, λ∈[0,1].
Citation Information
Patent Citations
Lightweight retinal vessel segmentation method based on attention mechanism
CN115760872A
Hybrid attention retinal vessel segmentation method based on residual U-shaped network
CN116363060A
Retinal vessel segmentation method based on dense connection adaptive receptive field
CN116385363A
Retinal vessel segmentation method fusing pixel connectivity and system therefor
WO2024108660A1
