Retinal vessel segmentation method and system based on improved U-Net

By introducing dense convolutional layer and SimAM attention mechanism in U-Net, the problems of detail feature loss and noise interference in traditional U-Net are solved, and the accuracy and robustness of retinal vascular segmentation are improved.

CN118941577APending Publication Date: 2024-11-12SHANGHAI INST OF TECH

Patent Information

Application Number
CN202410980330.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional U-Net and its variants will lose some detailed features during encoding and decoding, especially for small blood vessels in the retina. The segmentation effect is not ideal and is easily disturbed by noise, resulting in a decrease in segmentation accuracy.

Method used

Introduce dense convolutional layer and SimAM parameterless attention mechanism to map shallow and deep features through residual structures to enhance feature reuse, and weighted fusion of feature maps through SimAM attention mechanism to reduce gradient vanishing and overfitting problems .

Benefits of technology

The accuracy of retinal vascular segmentation is significantly improved, especially the segmentation effect in tiny blood vessels and complex backgrounds, effectively reducing the interference of background noise on segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004955869840000031
    Figure BDA0004955869840000031
  • Figure BDA0004955869840000061
    Figure BDA0004955869840000061
  • Figure FDA0004955869830000021
    Figure FDA0004955869830000021
Patent Text Reader

Abstract

The invention discloses a retinal vessel segmentation method and system based on improved U-Net, and the method comprises the steps: collecting a 3D fundus medical image, and carrying out the extraction and integration of image features through convolution, pooling and up-sampling operation through a symmetric coding-decoding structure; a dense convolutional layer and a SimAM parameter-free attention mechanism are introduced, shallow and deep features are mapped through a residual structure, and feature reusability is enhanced; gradient disappearance is reduced through a residual structure, a SimAM attention mechanism is used for carrying out weighted fusion on a feature map, and the feature representation capability of the model is improved. In combination with a SimAM attention mechanism, the model can more accurately pay attention to important features related to tasks, so that the precision of retinal blood vessel segmentation is remarkably improved, and especially the segmentation effect in small blood vessels and complex backgrounds is improved. According to the SimAM parameter-free attention mechanism, a feature similarity matrix is calculated, and weighted fusion is carried out on feature maps, so that the model can effectively reduce interference of background noise on blood vessel segmentation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and specifically to a retinal blood vessel segmentation method and system based on improved U-Net. Background Art

[0002] With the development of medical image processing technology, the diagnosis and treatment of ophthalmic diseases have also made significant progress. As a key step in ophthalmic image analysis, retinal vessel segmentation is crucial for evaluating fundus health, monitoring disease progression, and guiding treatment plans. Early retinal vessel segmentation mainly relied on manual methods, and doctors completed image segmentation through naked eye observation and manual annotation. However, this method is not only time-consuming and labor-intensive, but also highly subjective, easily affected by human factors, and difficult to ensure consistency and accuracy. With the development of computer vision technology, automatic segmentation methods based on traditional image processing algorithms have gradually been applied to retinal vessel segmentation, such as threshold segmentation, edge detection, and region growing. These methods have improved segmentation efficiency to a certain extent, but there are still problems such as sensitivity to noise and limited feature extraction capabilities.

[0003] In recent years, the rapid development of deep learning technology has brought new opportunities for medical image processing. In particular, the introduction of convolutional neural networks (CNNs) has significantly improved the accuracy and efficiency of retinal blood vessel segmentation. Among them, the U-Net model is widely used in medical image segmentation tasks because of its symmetrical encoding-decoding structure, which can extract deep semantic information while maintaining high-resolution features. However, traditional U-Net and its variants still face challenges in dealing with complex vascular structures and noise interference. First, the U-Net model loses some detailed features during the encoding and decoding process, especially for small blood vessels in the retina, and the segmentation effect is not ideal. Secondly, the existing deep learning algorithms are easily disturbed by background noise when dealing with noise, resulting in a decrease in segmentation accuracy. In addition, as the network depth increases, the gradient vanishing and overfitting problems during training also limit the further improvement of model performance. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the traditional U-Net and its variants will lose some detail features during the encoding and decoding process, especially for the tiny blood vessels in the retina, the segmentation effect is not ideal and these important details cannot be accurately extracted.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a retinal blood vessel segmentation method based on improved U-Net, comprising:

[0007] Collect 3D fundus medical images, use the symmetrical encoding-decoding structure, and extract and integrate image features through convolution, pooling, and upsampling operations;

[0008] Introducing dense convolutional layers and SimAM parameter-free attention mechanism, mapping shallow and deep features through residual structure, and enhancing feature reusability;

[0009] The residual structure is used to reduce gradient vanishing, and the SimAM attention mechanism is used to perform weighted fusion on feature maps to improve the model's feature representation capabilities.

[0010] As a preferred solution of the retinal blood vessel segmentation method based on improved U-Net described in the present invention, wherein: the symmetrical encoding-decoding structure includes an encoding part and a decoding part;

[0011] The encoding part is to expand the receptive field of the image through pooling and convolution operations and extract high-level abstract features;

[0012] The decoding part restores the resolution of the feature map through convolution and upsampling operations. Skip connections can connect and integrate feature maps of the same resolution to retain more detail information. Finally, the output of the U-Net model is converted into a binary segmented image through 1×1 convolution.

[0013] As a preferred solution of the retinal blood vessel segmentation method based on the improved U-Net described in the present invention, wherein: the dense convolution layer includes, the output of each feature extraction network is merged through the dense convolution layer, and the merged output will be used as the input of the next layer;

[0014] Each convolution operation in the dense convolution layer is a 3×3 convolution operation, followed by BN for data normalization and ReLU activation function for increasing the nonlinear characteristics of the network;

[0015] DropBlock is added as a regularization method in the dense convolution layer to prevent the network from overfitting. DropBlock prevents overfitting by randomly removing data blocks.

[0016] As a preferred solution of the retinal blood vessel segmentation method based on the improved U-Net described in the present invention, the dense convolution layer also includes the output of the current layer being fused with the output of all previous layers to form a new input feature, which is expressed as follows:

[0017] α l =CDBR([α 0 ,α 1 ,…,α l-1 ])

[0018] Among them, [α 0 ,α1 ,...,α l-1 ] represents the merging of feature maps from layer 0 to layer l-1, CDBR represents a layer of feature extraction network, C represents a 3×3 convolution operation, D represents the use of DropBlock to randomly remove data blocks to prevent overfitting, B represents BN for data standardization, and R represents the ReLU activation function;

[0019] In the feature extraction network, the number of channels of each convolution is set to 1 / 6, 1 / 3, and 1 / 2 of the total number of channels of the network, respectively, to ensure that the feature extraction efficiency is improved while maintaining a reasonable number of parameters.

[0020] As a preferred solution of the retinal blood vessel segmentation method based on the improved U-Net described in the present invention, wherein: mapping shallow and deep features through the residual structure includes, the residual network operation formula is expressed as:

[0021] h i+1 =h i +F(h i ,w i )

[0022] Among them, h i +1 represents the output mapping of the network layer l+1, h i represents the identity mapping of the ιth layer of the network, F(h l ,w l ) represents the residual mapping.

[0023] As a preferred solution of the retinal blood vessel segmentation method based on improved U-Net described in the present invention, wherein: the SimAM attention layer includes weighted fusion of information in the feature map by calculating the feature similarity matrix, retaining important features and filtering irrelevant information;

[0024] The feature similarity matrix of the SimAM attention mechanism reflects the importance of each feature and normalizes it into a probability distribution so that the model can focus more specifically on important features related to the task;

[0025] The SimAM attention layer is used in the deep residual enhancement module, combined with the dense convolution layer, to map shallow and deep features through the residual structure to enhance the reusability of features;

[0026] The feature normalization process formula of SimAM is expressed as:

[0027]

[0028] Among them, s(i,j) represents an element of the feature similarity matrix. By normalizing the feature similarity matrix into a probability distribution, the importance of each feature can be reflected.

[0029] As a preferred solution of the retinal blood vessel segmentation method based on improved U-Net described in the present invention, wherein: the deep residual enhancement module includes a dense convolution layer and a SimAM attention layer;

[0030] The output of each feature extraction network is combined through dense convolutional layers to solve the overfitting and gradient vanishing problems respectively;

[0031] The merged output is passed to the SimAM attention module to recalibrate the feature information, allowing the network to focus more on parsing the structural details of the blood vessels and performing pixel classification;

[0032] The original input is added to the current output after 1×1 convolution. This module combines the characteristics of ResNet and DenseNet, improves the network's ability in feature transmission, and enhances the effect of vascular image segmentation.

[0033] A retinal blood vessel segmentation system based on improved U-Net, wherein:

[0034] The U-Net model consists of an encoding part and a decoding part;

[0035] The encoding part involves extracting high-level abstract features through convolution and pooling operations.

[0036] The decoding part includes restoring the resolution of the feature map through convolution and upsampling operations, and using skip connections to retain more detailed information;

[0037] The deep residual enhancement module includes dense convolutional layers and SimAM attention layers. It maps shallow and deep features through residual structures, enhances feature reusability, and accelerates network convergence.

[0038] The dense convolution layer includes the output of the current layer, which is fused with the output of all previous layers. The 3×3 convolution operation is combined with BN and ReLU activation functions for feature extraction. The DropBlock regularization method prevents overfitting and randomly removes data blocks to enhance network robustness.

[0039] The SimAM attention mechanism involves calculating the feature similarity matrix and weighted fusion of the information in the feature map to improve the model's feature representation capabilities and effectively solve the problem of shallow feature redundancy.

[0040] The data preprocessing and enhancement module performs preprocessing through grayscale transformation, histogram equalization and gamma transformation to improve image contrast and detail information, and uses data enhancement methods to expand the data set and improve the generalization ability of the model.

[0041] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0042] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of any one of the methods of the present invention.

[0043] Beneficial effects of the present invention: The dense convolution layer introduced in the retinal vessel segmentation method based on the improved U-Net provided by the present invention improves the reusability and detail retention ability of features through dense connections, and effectively solves the problem of feature information loss caused by traditional convolution operations. Combined with the SimAM attention mechanism, the model can focus more accurately on important features related to the task, thereby significantly improving the accuracy of retinal vessel segmentation, especially the segmentation effect in small blood vessels and complex backgrounds. The SimAM parameter-free attention mechanism calculates the feature similarity matrix and performs weighted fusion on the feature map, so that the model can effectively reduce the interference of background noise on vessel segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0045] Figure 1 An overall flow chart of a retinal blood vessel segmentation method based on an improved U-Net provided in the first embodiment of the present invention;

[0046] Figure 2 A deep residual enhancement module diagram of a retinal blood vessel segmentation method based on an improved U-Net provided in the first embodiment of the present invention;

[0047] Figure 3 A SimAM structure diagram of a retinal blood vessel segmentation method based on an improved U-Net provided in the first embodiment of the present invention;

[0048] Figure 4 A network architecture diagram of a retinal blood vessel segmentation method based on an improved U-Net provided in the first embodiment of the present invention;

[0049] Figure 5 The segmentation results of each algorithm in a retinal vessel segmentation method based on an improved U-Net provided in the second embodiment of the present invention on the DRIVE dataset;

[0050] Figure 6 The segmentation results of various algorithms in a retinal vessel segmentation method based on an improved U-Net provided in the second embodiment of the present invention on the STARE dataset. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0052] Example 1, reference Figure 1 , as an embodiment of the present invention, provides a retinal blood vessel segmentation method based on an improved U-Net, comprising:

[0053] S1: Collect 3D fundus medical images and use a symmetrical encoding-decoding structure to extract and integrate image features through convolution, pooling, and upsampling operations.

[0054] The U-Net model uses a symmetrical encoding and decoding structure to extract and integrate image features through operations such as convolution, pooling, and upsampling. In the encoding part, the receptive field of the image is expanded through pooling and convolution operations, while the resolution of the input image is reduced, thereby extracting high-level abstract features. In the decoding part, the resolution of the feature map is restored through convolution and upsampling, so that the skip connection can connect and integrate feature maps of the same resolution, thereby retaining more detail information. Finally, the output of the model is converted into a binary segmentation image through 1×1 convolution. Two 3×3 convolutions and ReLU activation function operations are performed at each stage of the network, while other operations include 1×1 convolution, maximum pooling, and upsampling operations. Through these operations, after the input image is extracted and analyzed by the model, an accurate binary segmentation image can be obtained, effectively achieving the task of medical image segmentation. However, the network cannot extract some tiny blood vessels, and there are phenomena such as insufficient segmentation of microvessels and segmentation breaks at the intersection of blood vessels.

[0055] S2: Introducing dense convolutional layers and SimAM parameter-free attention mechanism, mapping shallow and deep features through residual structure, and enhancing feature reusability.

[0056] As the depth of the convolutional neural network increases, problems such as gradient vanishing and network overfitting may occur during the training process, thus affecting the segmentation effect of blood vessels. Inspired by the residual network and attention mechanism, this paper proposes a deep residual enhancement module, whose structure is as follows: Figure 1As shown in the figure, it mainly includes dense convolution layers and SimAM attention layers. The deep residual enhancement module maps shallow and deep features through the residual structure, enhances the reusability of features, and accelerates the convergence of the network, which helps to alleviate the gradient vanishing problem encountered during network propagation. The calculation formula of the residual network is as follows:

[0057] h i+1 =h i +F(h i ,w i )

[0058] Among them, h i +1 represents the output mapping of the network layer l+1, h i represents the identity mapping of the ιth layer of the network, F(h l ,w l ) represents the residual mapping.

[0059] At the same time, in order to solve the problem of feature redundancy in shallow features, the present invention introduces SimAM parameter-free attention in the deep residual enhancement module to perform weighted fusion of information at different positions in the feature map. The SimAM structure diagram is shown in Figure 2 The most significant advantage of this module is its powerful feature enhancement representation capability and efficient computing performance, which can effectively weight and filter feature maps without introducing additional parameters. The feature normalization process of SimAM can be expressed as follows:

[0060]

[0061] Among them, s(i,j) represents an element of the feature similarity matrix. By normalizing the feature similarity matrix into a probability distribution, the importance of each feature can be reflected.

[0062] This process helps the model focus more specifically on important features related to the task, thus highlighting the advantage of SimAM in enhancing the model's feature representation capabilities. The present invention uses the SimAM attention mechanism to effectively solve the problem of vascular information being confused by interference factors, and further improves the segmentation accuracy of retinal blood vessels.

[0063] S3: Reduce gradient vanishing through residual structure, use SimAM attention mechanism to perform weighted fusion on feature maps, and improve the model feature representation capability.

[0064] Multiple consecutive traditional convolutions in U-net may cause the loss of vascular features, resulting in incomplete segmentation of microvessels. Therefore, the present invention incorporates densely connected convolution layers in the module. In this structure, the output of the current layer is fused with the output of all previous layers, and then the fusion result is used as the input of the subsequent layer to improve the utilization of features, that is:

[0065] α l =CDBR([α 0 ,α 1 ,…,α l-1 ])

[0066] Among them, [α 0 ,α 1 ,…,α l-1 ] represents the merging of feature maps from layer 0 to layer l-1, CDBR represents a layer of feature extraction network, C represents a 3×3 convolution operation, D represents the use of DropBlock to randomly remove data blocks to prevent overfitting, B represents BN for data standardization, and R represents the ReLU activation function.

[0067] In the dense convolution layer, the number of channels for each convolution is set to 1 / 6, 1 / 3 and 1 / 2 of the total number of channels in the network, respectively, in order to improve the model's efficiency in extracting features while maintaining a reasonable amount of network parameters. In addition, the present invention adds DropBlock as a regularization method in the feature extraction network. Compared with the simple method of dropout that randomly discards activations to prevent overfitting of the neural network, DropBlock discards units in adjacent areas of the Featuremap together. This method performs better than dropout in the convolution layer and can effectively prevent network overfitting.

[0068] The core idea of ​​the deep residual enhancement module is as follows: first, the output of each feature extraction network is merged through a dense convolution layer to solve the overfitting and gradient vanishing problems respectively; then, the merged output is passed to the SimAM attention module to recalibrate the feature information, so that the network can focus more on analyzing the structural details of the blood vessels and performing pixel classification; finally, the original input is added to the current output after 1×1 convolution. This module combines the characteristics of ResNet and DenseNet, improves the network's ability in feature transmission, and enhances the effect of vascular image segmentation. Figure 3 The network architecture of the retinal vessel segmentation algorithm of the present invention is shown in FIG. 4 , which is based on a U-Net network and includes a 4-layer encoding-decoding structure.

[0069] On the other hand, this embodiment also provides a retinal blood vessel segmentation system based on an improved U-Net, which includes:

[0070] The U-Net model consists of an encoding part and a decoding part.

[0071] The encoding part involves extracting high-level abstract features through convolution and pooling operations.

[0072] The decoding part includes restoring the resolution of the feature map through convolution and upsampling operations, and using skip connections to retain more detailed information.

[0073] The deep residual enhancement module includes dense convolutional layers and SimAM attention layers, which maps shallow and deep features through residual structures, enhances feature reusability, and accelerates the convergence of the network.

[0074] The dense convolution layer includes the output of the current layer, which is fused with the output of all previous layers. A 3×3 convolution operation is used in combination with BN and ReLU activation functions for feature extraction. The DropBlock regularization method prevents overfitting and randomly removes data blocks to enhance network robustness.

[0075] The SimAM attention mechanism involves calculating the feature similarity matrix and weighted fusion of the information in the feature map to improve the model's feature representation capabilities and effectively solve the problem of shallow feature redundancy.

[0076] The data preprocessing and enhancement module performs preprocessing through grayscale transformation, histogram equalization and gamma transformation to improve image contrast and detail information, and uses data enhancement methods to expand the data set and improve the generalization ability of the model.

[0077] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0078] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0079] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0080] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0081] Example 2

[0082] Reference Figure 5-6, which is an embodiment of the present invention, model development and experiments are carried out on the Windows 10 operating system, and two public fundus image datasets, DRIVE and STARE, are used to evaluate the model performance. The model is trained for a total of 50 rounds, with 6 samples processed in each batch. Adam is used as the optimization method and cross entropy is used as the loss evaluation criterion. The initial learning rate is set to 0.001, and is adjusted through a polynomial decay strategy to accelerate model convergence. The experiment introduces an early stopping strategy. If the loss on the validation set does not improve within 15 rounds, in order to intuitively verify the effect of the retinal vascular segmentation algorithm proposed in the present invention, it is compared with three advanced algorithms: U-Net, AttU-Net and DenseU-Net. The segmentation results of each algorithm on the DRIVE and STARE datasets are shown in the figure. Figure 4-5 As shown, in comparison, the algorithm proposed in the present invention can accurately segment the main vessels and micro-vessels, especially in the areas where the vessels cross or are adjacent, showing fewer breaks and omissions. In summary, the present invention shows superior performance in retinal vessel segmentation compared to other methods.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A retinal blood vessel segmentation method based on improved U-Net, characterized in that: include: Collect 3D fundus medical images, use the symmetrical encoding-decoding structure, and extract and integrate image features through convolution, pooling, and upsampling operations; Introducing dense convolutional layers and SimAM parameter-free attention mechanism, mapping shallow and deep features through residual structure, and enhancing feature reusability; The residual structure is used to reduce gradient vanishing, and the SimAM attention mechanism is used to perform weighted fusion on feature maps to improve the model's feature representation capabilities.

2. The retinal blood vessel segmentation method based on improved U-Net according to claim 1, characterized in that: The symmetrical encoding-decoding structure includes an encoding part and a decoding part; The encoding part is to expand the receptive field of the image through pooling and convolution operations and extract high-level abstract features; The decoding part restores the resolution of the feature map through convolution and upsampling operations. Skip connections can connect and integrate feature maps of the same resolution to retain more detail information. Finally, the output of the U-Net model is converted into a binary segmented image through 1×1 convolution.

3. The retinal blood vessel segmentation method based on improved U-Net according to claim 2, characterized in that: The dense convolution layer includes: the output of each feature extraction network is merged through the dense convolution layer, and the merged output will be used as the input of the next layer; Each convolution operation in the dense convolution layer is a 3×3 convolution operation, followed by BN for data normalization and ReLU activation function for increasing the nonlinear characteristics of the network; DropBlock is added as a regularization method in the dense convolution layer to prevent the network from overfitting. DropBlock prevents overfitting by randomly removing data blocks.

4. The retinal blood vessel segmentation method based on improved U-Net according to claim 3, characterized in that: The dense convolution layer also includes the fusion of the output of the current layer with the output of all previous layers to form a new input feature, which is expressed as: a l =CDBR([a 0 ,a 1 ,…,a l-1 ]) Among them, [α 0 ,α 1 ,…,α l-1 ] represents the merging of feature maps from layer 0 to layer l-1, CDBR represents a layer of feature extraction network, C represents a 3×3 convolution operation, D represents the use of DropBlock to randomly remove data blocks to prevent overfitting, B represents BN for data standardization, and R represents the ReLU activation function; In the feature extraction network, the number of channels of each convolution is set to 1 / 6, 1 / 3, and 1 / 2 of the total number of channels of the network, respectively, to ensure that the feature extraction efficiency is improved while maintaining a reasonable number of parameters.

5. The retinal blood vessel segmentation method based on improved U-Net according to claim 4, characterized in that: Mapping shallow and deep features through the residual structure includes the residual network operation formula as follows: h i+1 =h i +F(h i ,w i ) Among them, h i +1 represents the output mapping of the network layer l+1, h i represents the identity mapping of the ιth layer of the network, F(h l ,w l ) represents the residual mapping.

6. The retinal blood vessel segmentation method based on improved U-Net according to claim 5, characterized in that: The SimAM attention layer includes,weighted fusion of information in feature maps by calculating feature similarity,matrix, retaining important features and filtering irrelevant information; The feature similarity matrix of the SimAM attention mechanism reflects the importance of each feature and normalizes it into a probability distribution so that the model can focus more specifically on important features related to the task; The SimAM attention layer is used in the deep residual enhancement module, combined with the dense convolution layer, to map shallow and deep features through the residual structure to enhance the reusability of features; The feature normalization process formula of SimAM is expressed as: Among them, s(i,j) represents an element of the feature similarity matrix. By normalizing the feature similarity matrix into a probability distribution, the importance of each feature can be reflected.

7. The retinal blood vessel segmentation method based on improved U-Net according to claim 6, characterized in that: The deep residual enhancement module includes dense convolutional layers and SimAM attention layers; The output of each feature extraction network is combined through dense convolutional layers to solve the overfitting and gradient vanishing problems respectively; The merged output is passed to the SimAM attention module to recalibrate the feature information, allowing the network to focus more on parsing the structural details of the blood vessels and performing pixel classification; The original input is added to the current output after 1×1 convolution. This module combines the characteristics of ResNet and DenseNet, improves the network's ability in feature transmission, and enhances the effect of vascular image segmentation.

8. A retinal blood vessel segmentation system based on an improved U-Net using the method according to any one of claims 1 to 7, characterized in that: The U-Net model consists of an encoding part and a decoding part; The encoding part includes extracting high-level abstract features through convolution and pooling operations; The decoding part includes restoring the resolution of the feature map through convolution and upsampling operations, and using skip connections to retain more detailed information; The deep residual enhancement module includes dense convolutional layers and SimAM attention layers. It maps shallow and deep features through residual structures, enhances feature reusability, and accelerates network convergence. The dense convolution layer includes the output of the current layer, which is fused with the output of all previous layers. The 3×3 convolution operation is combined with BN and ReLU activation functions for feature extraction. The DropBlock regularization method prevents overfitting and randomly removes data blocks to enhance network robustness. The SimAM attention mechanism involves calculating the feature similarity matrix and weighted fusion of the information in the feature map to improve the model's feature representation capabilities and effectively solve the problem of shallow feature redundancy. The data preprocessing and enhancement module performs preprocessing through grayscale transformation, histogram equalization and gamma transformation to improve image contrast and detail information, and uses data enhancement methods to expand the data set and improve the generalization ability of the model.

9. A computer device comprising: Memory and processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the retinal blood vessel segmentation method based on the improved U-Net are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a retinal blood vessel segmentation method based on an improved U-Net are implemented.

Citation Information

Patent Citations

  • Retinal blood vessel image segmentation method and device and computer equipment

    CN114283158A

  • Retina image blood vessel segmentation method based on improved U-Net network

    CN114881962A

  • Image segmentation method and device, electronic equipment and computer storage medium

    CN114972363A

  • Medical image segmentation method based on u-net

    US20220309674A1

Cited By

  • Method for improving imaging quality of Compton camera

    CN119444907A

  • Automatic intima-media thickness measuring method for blood vessel diameter detection based on ultrasonic image

    CN120563592A