Retina vessel segmentation method based on difference feature fusion

Abstract

The invention relates to a retinal vessel segmentation method based on difference feature fusion, and belongs to the field of medical image segmentation. The invention provides a lightweight retinal vessel segmentation network Diffuse-Net based on differential feature fusion, and the lightweight retinal vessel segmentation network Diffuse-Net is used for accurate and efficient eye vessel image segmentation. On the basis of the fact that an eye blood vessel image has obvious local feature difference, a hybrid differential convolution module HyDC is provided, and the HyDC deploys common convolution and four different differential convolves in parallel, so that various direction features and edge information of eye blood vessels are comprehensively captured, and the contrast ratio between the blood vessels and the background is enhanced; besides, an integrated feature fusion module IFFM is further designed to effectively integrate multi-scale information, and the IFFM integrates high-level features and low-level features of different scales on the basis of grouping aggregation, so that robust feature representation is realized. The method disclosed by the invention obtains leading performance in an eye vascular structure segmentation task.

Description

Technical Field

[0001] The present invention belongs to the field of medical image segmentation, and specifically relates to a retinal blood vessel segmentation method based on difference feature fusion, which is used to accurately segment ocular blood vessel structures with complex morphology and diverse features; it has a wide range of applications in ophthalmic diagnosis and treatment, such as early detection of diabetic retinopathy. Background Art

[0002] Accurate segmentation of ocular blood vessels is crucial for the early diagnosis and monitoring of a variety of ophthalmic diseases. Morphological and structural changes in retinal blood vessels are early symptoms of ocular and systemic diseases such as diabetic retinopathy, hypertensive retinopathy, and glaucoma. Existing blood vessel segmentation methods mainly include traditional image processing methods and deep learning methods that have emerged in recent years. Although traditional methods are computationally efficient, they perform poorly in complex backgrounds and low-contrast images. Deep learning methods such as U-Net and its variants have significantly improved segmentation accuracy, but because they use ordinary convolution to extract features, they ignore the local feature differences of blood vessels and are difficult to adapt to blood vessels of different thicknesses and curvatures. They may not be flexible enough for complex blood vessel intersections and branching structures.

[0003] In order to solve the problem that ordinary convolution cannot accurately segment ocular blood vessels, several innovative convolution methods and network structures have emerged recently. These new methods aim to improve the accuracy of ocular blood vessel segmentation. For example, Lv et al. [1] proposed a U-Net with attention guidance and dilated convolution. They used the attention mechanism to regress the bounding box of the retinal region to guide the model to pay more attention to the vascular region. At the same time, the dilated convolution was used to replace the ordinary convolution in the feature layer to expand the receptive field and reduce the amount of computation. MSCNN-AM [2] is a multi-scale convolutional neural network with an attention mechanism. It introduces dilated separable convolutions with different dilation rates to extract blood vessels of different scales, thereby better capturing global and multi-scale vascular information. Arsalan et al. [3] proposed a hint block that combines asymmetric kernel convolution, depthwise separable convolution and ordinary convolution to improve the performance of the network in order to solve the problem that the existing encoder-decoder architecture often ignores the context of retinal blood vessel structure. Yang et al. [4] proposed an improved U-net based on Gabor wavelet and Gabor convolution network, which improved the performance of retinal blood vessel segmentation by enhancing the recognition ability of small and low-contrast blood vessels.

[0004] like Figure 1As shown in the figure, mainstream methods use a variety of innovative convolution techniques, including dynamic convolution, dilated convolution, depthwise separable convolution and Gabor convolution, or cleverly combine these techniques. Dynamic convolution can adaptively adjust the convolution kernel according to the input, enhancing the flexibility of the model and its adaptability to complex patterns, but it may increase the complexity and computational overhead of the model [5]; dilated convolution captures multi-scale information by expanding the receptive field, and can effectively identify vascular structures of different scales. However, while expanding the receptive field, it may introduce the gradient vanishing problem, affecting the training stability of the model [6]; depthwise separable convolution significantly reduces the number of parameters and computational cost by decomposing the standard convolution into depthwise convolution and pointwise convolution, and improves the running efficiency of the model. Although this method may limit the ability to express features, especially when processing highly complex images [7]; Gabor convolution [8] uses predefined filter parameters to extract directional and texture features, enhancing the sensitivity to specific textures and edges, and helping to identify subtle vascular structures, but it relies on fixed filter parameters, lacks flexibility, and is difficult to adapt to changing image features. In addition, these convolution techniques often have difficulty in effectively capturing multi-directional and multi-scale feature information at the same time, which limits their performance in complex vascular structure segmentation tasks. Therefore, although these innovative convolution methods have improved the performance of the model to a certain extent, they still have certain limitations when processing complex and changeable retinal vascular images, and need to be further combined with other technical means to achieve higher segmentation accuracy and robustness. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and provide a retinal blood vessel segmentation method based on difference feature fusion to achieve accurate and efficient segmentation of retinal blood vessels.

[0006] To achieve the above object, the technical solution of the present invention is: a retinal blood vessel segmentation method based on difference feature fusion, comprising:

[0007] A lightweight retinal vessel segmentation network Diffuse-Net based on differential feature fusion is proposed for accurate and efficient retinal vessel image segmentation; Diffuse-Net includes feature extraction stage and feature fusion stage;

[0008] In the feature extraction stage, a hybrid differential convolution module HyDC is proposed to comprehensively capture various directional features and edge information of retinal blood vessels and enhance the contrast between blood vessels and background.

[0009] In the feature fusion stage, an integrated feature fusion module IFFM is designed to integrate high-level features and low-level features of different scales on the basis of grouping aggregation to achieve robust feature representation.

[0010] In one embodiment of the present invention, Diffuse-Net adopts an encoder-decoder structure.

[0011] In one embodiment of the present invention, the hybrid differential convolution module HyDC is composed of parallel ordinary convolutions and four differential convolutions. Four different types of differential convolutions are used to enhance gradient level information, and a conventional convolution is introduced to capture intensity level information. After obtaining these five different convolution weights, these prior information is provided to subsequent conventional convolution layers.

[0012] In one embodiment of the present invention, the four differential convolutions include center differential convolution CDC, angle differential convolution ADC, horizontal differential convolution HDC and vertical differential convolution VDC.

[0013] In one embodiment of the present invention, the central difference convolution CDC is specifically implemented as follows:

[0014] In the central difference convolution, a standard 3×3 convolution kernel is first used to obtain normal weights Then, the dimensions of the obtained standard convolution weights are rearranged; after that, a new weight tensor W is created for the central difference convolution cd , and copy the normal weight W to W cd middle;

[0015] For the non-central position of the center difference convolution kernel, set:

[0016] W(c in ,c out ,k1,k2)→W(c in ,c out ,k1×k2)

[0017] W cd (c in ,c out ,k1×k2)=W(c in ,c out ,k1×k2)

[0018] Among them, c in is the number of input channels, c out is the number of output channels, k1 and k2 are the height and width of the convolution kernel respectively;

[0019] For the center position of the central difference convolution kernel, its weight is adjusted by subtracting the sum of the weights of all other positions; this step realizes the key transformation from standard convolution to central difference convolution:

[0020]

[0021] Among them, i is an index variable used to indicate the position of each element in the 3×3 convolution kernel;

[0022] Finally, the resized weights are rearranged back to the dimensions of the original convolutional weights.

[0023] In one embodiment of the present invention, the angle differential convolution ADC is specifically implemented as follows:

[0024] In the angular differential convolution, a 3×3 convolution kernel is first used to obtain the initial weights and rearrange its dimensions. An additional parameter θ is set to control the strength of the difference so that each position of the convolution weight is adjusted according to θ. The following formula is used to express the process of obtaining the angular differential convolution weights:

[0025] W ad =(c in ,c out ,i)=W(c in ,c out ,i)-θ·W(c in ,c out ,j(i))

[0026] Here, θ is a parameter that controls the differential intensity and j(i) is a function that defines the angular offset with respect to position:

[0027]

[0028] mod refers to the remainder operation; for each position i, the weight is adjusted according to the corresponding j(i).

[0029] In one embodiment of the present invention, the horizontal differential convolution HDC and the vertical differential convolution VDC are specifically implemented as follows:

[0030] In the horizontal differential convolution, we first use a 3×3 convolution kernel to obtain the standard weights, and create a new tensor of the same shape to store the weights of the horizontal differential convolution. We obtain W by equivalently rearranging the standard weights. hd Specifically, the value of the first column of the standard convolution kernel weight is assigned to the first column of the horizontal difference convolution weight, and the negative value of the third column of the standard convolution kernel weight is assigned to the third column of the horizontal difference convolution weight, and the weight of the horizontal difference convolution kernel is derived. hd :

[0031]

[0032] The weights of the vertical difference convolution kernel can be derived similarly by assigning the positive values ​​of the first row and the negative values ​​of the third row of the standard convolution kernel weights to the first and third rows of the vertical difference convolution, respectively:

[0033]

[0034] In one embodiment of the present invention, the integrated feature fusion module IFFM is specifically implemented as follows:

[0035] The two inputs of the integrated feature fusion module IFFM are: low-level features X l and high-level features X h ;

[0036] First, use 1×1 convolution and bilinear interpolation to adjust the high-level features X h The number of channels and dimensions are adjusted to obtain the feature X' h , so that it is consistent with the low-level feature X l With the same number of channels and spatial dimensions:

[0037] X′ h =Bilinear(Conv 1×1 (X h ))

[0038] Among them, Bilinear represents the bilinear interpolation operation, Conv 1×1 Represents a 1×1 convolution operation;

[0039] Then, along the channel dimension, X' h and X l are divided into four sub-tensors; specifically, X' h is divided into four sub-tensors: x' h [0], x' h [1] x' h [2] and x' h [3], X l is divided into four sub-tensors: x l [0], x l [1] x l [2] and x l [3]:

[0040] x′ h [0],x′ h [1],x′ h [2],x′ h [3] = Split(X′ h ,4,dim=1)

[0041] x l [0],x l [1],x l [2],x l [3]=Split(X l ,4,dim=1)

[0042] Among them, Split refers to the chunk function in the Pytorch library; the dim parameter specifies the dimension of the operation, and dim=1 means operating along the channel dimension;

[0043] Subsequently, the four sub-tensors obtained from the low-level features and high-level features are matched and concatenated along the channel dimension; this process generates four sets of fused features, each containing information from both low-level and high-level features:

[0044] x′ k =concat(x′ h [k],x l [k]),k∈0,1,2,3

[0045] Among them, concat(·) represents the concatenation operation along the channel dimension, i.e., the first dimension;

[0046] Next, layer normalization is performed on these four sets of fused features, and dilated convolutions with 3×3 convolution kernels and different dilation rates are applied to extract information at different scales, resulting in four sets of information: x0, x1, x2, and x3:

[0047] x k =Conv d(3×3) (LN(x′ i )),i∈0,1,2,3

[0048] Here, d represents the dilation rate, LN refers to layer normalization, Conv d(3×3) Represents a 3×3 convolution operation;

[0049] Finally, the four sets of information obtained are concatenated, and then layer normalization and 1×1 standard convolution are performed to obtain the final feature X:

[0050] X=Conv 1×1 (LN(concat(x0,x1,x2,x3))).

[0051] In one embodiment of the present invention, the method uses a binary cross entropy-Dice loss function BCEDiceLoss to optimize the network; BCEDiceLoss is a composite loss function that combines binary cross entropy loss BCE with Dice loss; wherein,

[0052] Binary cross entropy loss is used to measure the error between the predicted probability and the true label of each pixel for binary classification problems; its formula is:

[0053]

[0054] Where N is the total number of pixels, y iis the true label of pixel i, which is 0 or 1, p i is the predicted probability of the i-th pixel;

[0055] Dice loss measures the similarity between the prediction and the true label by directly optimizing the Dice coefficient, targeting the problem of class imbalance, especially small area foreground segmentation; its formula is:

[0056]

[0057] BCEDiceLoss is a combined loss function that is a weighted sum of BCE loss and Dice loss. Its formula is:

[0058] L BCEDice =α·L BCE +β·L Dice

[0059] Among them, α and β are adjustable weight parameters used to balance the contribution of BCE loss and Dice loss in the total loss.

[0060] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0061] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a lightweight retinal vascular segmentation network DiFFuse-Net based on differential feature fusion, which is used for accurate and efficient segmentation of ocular vascular images. Based on the fact that ocular vascular images have obvious local feature differences, the present invention proposes a hybrid differential convolution module (HyDC). HyDC deploys ordinary convolution and four different differential convolutions in parallel, so as to comprehensively capture various directional features and edge information of ocular blood vessels and enhance the contrast between blood vessels and background. In addition, the present invention also designs an integrated feature fusion module (IFFM) to effectively integrate multi-scale information. On the basis of group aggregation, IFFM simply integrates high-level and low-level features of different scales to achieve robust feature representation; the method of the present invention has achieved leading performance in the task of ocular vascular structure segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Diagram of methods for extracting image features for different convolutions.

[0063] Figure 2 This is the network model architecture of the method of the present invention.

[0064] Figure 3 Comparison diagram of the segmentation results of each comparison method on three datasets. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0066] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0067] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0068] The present invention provides a retinal blood vessel segmentation method based on difference feature fusion, comprising:

[0069] A lightweight retinal vessel segmentation network Diffuse-Net based on differential feature fusion is proposed for accurate and efficient retinal vessel image segmentation; Diffuse-Net includes feature extraction stage and feature fusion stage;

[0070] In the feature extraction stage, a hybrid differential convolution module HyDC is proposed to comprehensively capture various directional features and edge information of retinal blood vessels and enhance the contrast between blood vessels and background.

[0071] In the feature fusion stage, an integrated feature fusion module IFFM is designed to integrate high-level features and low-level features of different scales on the basis of grouping aggregation to achieve robust feature representation.

[0072] The following is the specific implementation process of the present invention.

[0073] 1. Feature extraction and fusion

[0074] In order to enhance the expressive power and flexibility of features and reduce the amount of calculation, the method of the present invention proposes a new segmentation framework. This framework significantly improves the network's understanding and processing capabilities of ocular vascular images by introducing innovative feature extraction modules and feature fusion modules. Specifically, the feature extraction module adopts an improved differential convolution structure, which can more efficiently capture subtle vascular edge information while maintaining low computational complexity. In addition, in order to effectively fuse multi-scale features at different levels in the network, the present invention designs a novel integrated feature fusion module. This module is based on grouped convolution and integrates features of different scales in a concise manner to obtain rich global contextual information. The overall framework not only improves the accuracy of segmentation and the ability to restore details, but also optimizes computational efficiency, and is suitable for the needs of real-time ocular vascular image processing.

[0075] 2. Network structure design

[0076] like Figure 2 As shown, the algorithm adopts a typical encoder-decoder structure, which includes three stages: 1) convolution, 2) tokenized multilayer perceptron, and 3) integrated feature fusion. First, the present invention uses a hybrid differential convolution module to extract low-dimensional features from the input image in the first three stages of the encoder. In the hybrid differential convolution module, differential convolution is used to capture local feature differences in retinal images and enhance the representation and generalization capabilities of the model. Then, the tokenized multilayer perceptron module tokenizes and projects the input encoder features, and models the representation using a multilayer perceptron. Next, the features enter the decoder to reconstruct the segmentation results of the original image. The present invention introduces an integrated feature fusion module at each jump connection in the convolution stage to achieve effective feature fusion between the corresponding encoder and decoder layers, further integrating feature information from different stages, thereby enhancing the representation ability of the network.

[0077] 3. Hybrid differential convolution module

[0078] like Figure 2 As shown on the right, the Hybrid Differential Convolution Block (HyDC) consists of five different types of convolutions: standard convolution and four differential convolutions. The present invention deploys these five different convolutions in parallel for feature extraction. In the HyDC block, in addition to the common center differential convolution (CDC) and angle differential convolution (ADC), the present invention also introduces horizontal differential convolution (HDC) and vertical differential convolution (VDC) to integrate traditional local descriptors into the convolution layer [9].

[0079] In the center difference convolution, the present invention first uses the standard 3×3 convolution kernel to obtain the common weights Then, the present invention rearranges the dimensions of the obtained standard convolution weights. After that, the present invention creates a new weight tensor W for the central difference convolution cd , and copy the normal weights W to the new tensor W cd For the non-central position of the central difference convolution kernel, the present invention sets:

[0080] W(c in ,c out ,k1,k2)→W(c in ,c out ,k1×k2)

[0081] W cd (c in ,c out ,k1×k2)=W(c in ,c out ,k1×k2)

[0082] Among them, c in is the number of input channels, c out is the number of output channels, k1 and k2 are the height and width of the convolution kernel respectively.

[0083] For the center position of the center difference convolution kernel, the present invention adjusts its weight by subtracting the sum of the weights of all other positions. This step realizes the key transformation from standard convolution to center difference convolution:

[0084]

[0085] Wherein, i is an index variable, which is used to indicate the position of each element in the 3×3 convolution kernel. Finally, the present invention rearranges the adjusted weights back to the dimension of the original convolution weights.

[0086] Similarly, in order to obtain the weights of the angular differential convolution, the present invention still first uses a 3×3 convolution kernel to obtain the initial weights and rearranges its dimensions. The key difference is that the present invention requires an additional parameter θ to control the strength of the difference so that each position of the convolution weight is adjusted according to θ. The present invention can use the following formula to represent the process of obtaining the angular differential convolution weights:

[0087] W ad =(c in ,c out ,i)=W(c in ,c out ,i)-θ·W(c in ,c out ,j(i))

[0088] Among them, θ is a parameter that controls the differential intensity, i is still used to represent the index position of each element position in the 3×3 convolution kernel, and j(i) is a function used to define the angular offset related to the position:

[0089]

[0090] Here mod refers to the remainder operation. In this way, for each position i, the weight will be adjusted according to the corresponding j(i).

[0091] In the horizontal differential convolution, the present invention first uses a 3×3 convolution kernel to obtain the standard weights and creates a new tensor of the same shape to store the weights of the horizontal differential convolution. The present invention obtains W by equivalently rearranging the standard weights. hd Specifically, the present invention assigns the value of the first column of the standard convolution kernel weights to the first column of the horizontal differential convolution weights, and assigns the negative value of the third column of the standard convolution kernel weights to the third column of the horizontal differential convolution weights:

[0092]

[0093] It is worth noting that the form of this equivalent convolution kernel is similar to the traditional local descriptor. The standard horizontal difference convolution kernel can be regarded as a special case of the equivalent convolution kernel

[10] .

[0094] By assigning the positive values ​​of the first row and the negative values ​​of the third row of the standard convolution kernel weights to the first row and the third row of the vertical difference convolution, respectively, the present invention can similarly derive the weights of the vertical difference convolution kernel:

[0095]

[0096] The present invention uses four different types of differential convolution to enhance gradient level information and introduces a regular convolution to capture intensity level information. After obtaining these five different convolution weights, the present invention provides this prior information to the subsequent regular convolution layer. This method helps the subsequent convolution layer better capture the details and edge information in the image, thereby obtaining a more robust feature representation.

[0097] 4. Integrated feature fusion module

[0098] The structure of the integrated feature fusion module is as follows: Figure 2 As shown on the left. This module requires two inputs: low-level features X l and high-level features X h First, the present invention uses 1×1 convolution and bilinear interpolation to adjust the high-level features X h The number of channels and dimensions are adjusted to obtain the feature X' h , so that it is consistent with the low-level feature X lWith the same number of channels and spatial dimensions:

[0099] X′ h =Bilinear(Conv 1×1 (X h ))

[0100] Among them, Bilinear represents the bilinear interpolation operation, Conv 1×1 represents a 1×1 convolution operation. Then, the present invention transforms X' along the channel dimension h and Xl are divided into four sub-tensors. Specifically, X' h is divided into the following four sub-tensors: x' h [0], x' h [1] x' h [2] and x' h [3], while X l It is divided into x l [0], x l [1] x l [2] and x l [3]:

[0101] x′ h [0],x′ h [1],x′ h [2],x′ h [3] = Split(X′ h ,4,dim=1)

[0102] x l [0],x l [1],x l [2],x l [3]=Split(X l ,4,dim=1)

[0103] Among them, Split refers to the chunk function in the Pytorch library; the dim parameter specifies the dimension of the operation, and dim=1 means operating along the channel dimension.

[0104] Subsequently, the invention matches the four sub-tensors obtained from the low-level and high-level features and concatenates them along the channel dimension. This process generates four sets of fused features, each containing information from both low-level and high-level features:

[0105] x′ k =concat(x′ h [k],x l [k]),k∈0,1,2,3

[0106] Where concat(·) represents the concatenation operation along the channel dimension (i.e., the first dimension). Next, the present invention performs layer normalization on the four sets of fused features and applies dilated convolutions with 3×3 convolution kernels and different dilation rates to extract information of different scales, obtaining four sets of information: x0, x1, x2, and x3:

[0107] x k =Conv d(3×3) (LN(x′ i )),i∈0,1,2,3

[0108] Here, d represents the dilation ratio and LN refers to layer normalization.

[0109] Finally, the present invention connects the four sets of information obtained, and then performs layer normalization and 1×1 standard convolution to obtain the final feature X:

[0110] X=Conv 1×1 (LN(concat(x0,x1,x2,x3)))

[0111] This is done to achieve interaction between features of different scales.

[0112] 5. Training objectives

[0113] The present invention uses the binary cross entropy-Dice loss function (BCEDiceLoss) to optimize the network. BCEDiceLoss is a composite loss function that combines binary cross entropy loss

[11] (Binary Cross-Entropy Loss, BCE) with Dice loss

[12] (Dice Loss), and is widely used in binary segmentation tasks of medical images. This combination can simultaneously utilize the pixel-level classification capability of BCE and the class imbalance processing capability of Dice loss, thereby obtaining better segmentation performance in various scenarios.

[0114] Binary cross entropy loss is used to measure the error between the predicted probability and the true label of each pixel, and is usually used for binary classification problems. Its formula is:

[0115]

[0116] Where N is the total number of pixels, y i is the true label (0 or 1) of pixel i, p i is the predicted probability of the ith pixel.

[0117] Dice loss is specifically designed for class imbalance problems, especially small area foreground segmentation, by directly optimizing the Dice coefficient (which measures the similarity between the predicted and true labels). Its formula is:

[0118]

[0119] BCEDiceLoss is a combined loss function that is a weighted sum of BCE loss and Dice loss. Its formula is:

[0120] L BCEDice =α·L BCE +β·L Dice

[0121] Among them, α and β are adjustable weight parameters used to balance the contribution of BCE loss and Dice loss in the total loss. The common setting is α = 1 and β = 1, that is, the two are used in balance.

[0122] 6. Experiment

[0123] The model of the present invention is evaluated on three public retinal vascular image datasets. The DRIVE

[13] dataset contains 40 color images (565×584 pixels), which are evenly divided into training and test sets, and are accompanied by manual segmentation annotations. The CHASE_DB1

[14] dataset contains 28 color images (999×960 pixels), of which the present invention uses 20 for training and 8 for testing, and uses the first of the two annotation sets as the ground truth. The STARE

[15] dataset includes 20 retinal camera images (700×605 pixels), of which 16 are used for training and 4 are used for testing.

[0124] 6.1 Parameter Selection

[0125] The present invention resizes the images to 512×512 pixels and uses 80% of the dataset for training and 20% for testing. The present invention's method uses PyTorch 2.3.1 on a machine equipped with a 10-core Intel Xeon CPU, 128GB of memory, and an Nvidia RTX3090 GPU. An initial learning rate of 0.001 is used during training, and cosine annealing, Adam optimizer (weight decay 1e-4), and BCEDiceLoss are used.

[0126] The DRIVE dataset is a widely accepted standard in the field of retinal vessel segmentation. It provides official training and test sets, which makes it very suitable for standardized evaluation. The present invention conducts ablation experiments on this dataset to evaluate the effectiveness of each module of DiFFuse-Net. As shown in Table 1, HyDC and IFFM significantly improve the performance of the network in the retinal vessel segmentation task.

[0127] Table 1 Ablation experiment

[0128]

[0129] 6.2 Qualitative Comparison

[0130] Figure 3 The test results of various comparison methods and the DiFFuse-Net proposed in this invention on the DRIVE, CHASE_DB1 and STARE datasets are shown, and the corresponding true labels are provided by this invention. Compared with other methods, DiFFuse-Net can effectively preserve the fine structure of blood vessels during the segmentation process, including small branches and ends, which are usually easily ignored by traditional segmentation methods.

[0131] 6.3 Quantitative comparison

[0132] Comprehensive experiments were conducted on three widely used retinal vascular image datasets to evaluate the performance of the proposed method compared with other vascular segmentation methods. The evaluation focused on the following key indicators: accuracy, mean intersection over union (mIoU), F1 score, precision, and recall. Tables 2 to 4 show the comparison results between the proposed method and U-Net

[16] , DeepLabV3+

[17] , DS-Transunet

[18] , UNeXt

[19] , and DeepCrackAT

[20] on each dataset. The experimental results show that the proposed method outperforms the existing methods in multiple important indicators, demonstrating its excellent performance in the retinal vessel segmentation task.

[0133] Table 2 Results of various comparison methods on the DRIVE dataset

[0134]

[0135] Table 3 Results of various comparison methods on the CHASE_DB1 dataset

[0136]

[0137] Table 4 Results of various comparison methods on the STARE dataset

[0138]

[0139] Specifically, on the DRIVE and STARE datasets, the method of the present invention achieved the best performance in terms of accuracy, mIoU and F1 score. In the DRIVE dataset, the accuracy of the present invention reached 96.34%, which is 0.21 percentage points higher than the second best method UNeXt; the mIoU and F1 scores reached 65.94% and 79.23%, respectively, which are 1.17 and 0.64 percentage points higher than the second best method U-Net. In the STARE dataset, the accuracy, mIoU and F1 scores of the present invention reached 97.25%, 63.07% and 77.43%, respectively, which are 0.02, 0.97 and 0.9 percentage points higher than the second best method UNeXt.

[0140] In the CHASE_DB1 dataset, the proposed method achieved the best results in mIoU and F1 scores, which were 64.38% and 78.31% respectively, and the accuracy rate reached 97.02%, which was only 0.02 percentage points lower than the best method UNeXt, achieving the second best result. This shows that the overall performance of the proposed method on this dataset is still satisfactory, further verifying its generalization ability and robustness on different retinal vascular images.

[0141] In summary, the segmentation method proposed in the present invention has shown superior performance on multiple public datasets, significantly improving the accuracy of blood vessel segmentation and the ability to restore details, proving its potential and effectiveness in practical applications.

[0142] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0143] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0144] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0147] References:

[0148] [1]Lv Y,Ma H,Li J,et al.Attention guided U-Net with atrousconvolution for accurate retinal vessels segmentation[J].IEEEAccess,2020,8:32826-32839.

[0149] [2]Fu Q, Li S, Wang

[0150] [3]Arsalan M,Khan T M,Naqvi S S,et al.Prompt deep light-weight vesselsegmentation network(PLVS-Net)[J].IEEE / ACM Transactions on ComputationalBiology and Bioinformatics,2022,20(2):1363-1371.

[0151] [4]Yang C,Tang Y,Peng H,et al.RAGE-Net:Enhanced retinal vesselsegmentation U-shaped network using Gabor convolution[J].Digital SignalProcessing,2024:104643.

[0152] [5]Su Z,Fang L,Kang W,et al.Dynamic group convolution foraccelerating convolutional neural networks[C] / / Computer Vision–ECCV 2020:16thEuropean Conference,Glasgow,UK,August23–28,2020,Proceedings,Part VI16.Springer International Publishing,2020:138-155.

[0153] [6]Wang Y,Wang G,Chen C,et al.Multi-scaledilated convolution ofconvolutional neural network for image denoising[J].Multimedia ToolsandApplications,2019,78:19945-19960.

[0154] [7]Kaiser L,Gomez A N,Chollet F.Depthwise separable convolutions forneural machine translation[J].arXiv preprint arXiv:1706.03059,2017.

[0155] [8]Yuan Y,Zhang J,Wang Q.Deep Gabor convolution network for personre-identification[J].Neurocomputing,2020,378:387-398.

[0156] [9]Chen Z,He Z,Lu Z M.DEA-Net:Single image dehazing based on detail-enhancedconvolution and content-guided attention[J].IEEE Transactions onImage Processing,2024.

[0157]

[10] Li Q,Wang Q,Li X.Exploring the relationship between 2D / 3Dconvolution forhyperspectral image super-resolution[J].IEEE Transactions onGeoscience and Remote Sensing,2021,59(10):8693-8703.

[0158]

[11] Ruby U,Yendapalli V.Binary cross entropy with deep learningtechnique for imageclassification[J].Int.J.Adv.Trends Comput.Sci.Eng,2020,9(10).

[0159]

[12] Dice L R.Measures ofthe amount ofecologic association betweenspecies[J].Ecology,1945,26(3):297-302.

[0160]

[13] Staal J,AbràmoffM D,Niemeijer M,et al.Ridge-based vesselsegmentation in color imagesofthe retina[J].IEEE transactions on medicalimaging,2004,23(4):501-509.

[0161]

[14] Fraz M M,Remagnino P,Hoppe A,et al.An ensemble classification-based approach appliedto retinal blood vessel segmentation[J].IEEETransactions on Biomedical Engineering,2012,59(9):2538-2548.

[0162]

[15] Hoover A D,Kouznetsova V,Goldbaum M.Locating blood vessels inretinal images bypiecewise threshold probing of a matched filter response[J].IEEE Transactions on Medical imaging,2000,19(3):203-210.

[0163]

[16] Ronneberger O,Fischer P,Brox T.U-net:Convolutional networks forbiomedical imagesegmentation[C] / / Medical image computing and computer-assisted intervention–MICCAI 2015:18thinternational conference,Munich,Germany,October 5-9,2015,proceedings,part III 18.SpringerInternationalPublishing,2015:234-241.

[0164]

[17] Chen L C,Zhu Y,Papandreou G,et al.Encoder-decoder with atrousseparable convolutionfor semantic image segmentation[C] / / Proceedings of theEuropean conference on computer vision(ECCV).2018:801-818.

[0165]

[18] Lin A,Chen B,Xu J,et al.Ds-transunet:Dual swin transformer u-netfor medical imagesegmentation[J].IEEE Transactions on Instrumentation andMeasurement,2022,71:1-15.

[0166]

[19] Valanarasu J M J,Patel V M.Unext:Mlp-based rapid medical imagesegmentationnetwork[C] / / International conference on medical image computingand computer-assisted intervention.Cham:Springer Nature Switzerland,2022:23-33.

[0167]

[20] Lin Q,Li W,Zheng X,et al.DeepCrackAT:An effective cracksegmentation framework based on learning multi-scale crack features[J].Engineering Applications of Artificial Intelligence,2023,126:106876.。

[0168] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A retinal blood vessel segmentation method based on difference feature fusion, characterized in that: include: A lightweight retinal vessel segmentation network Diffuse-Net based on differential feature fusion is proposed for accurate and efficient retinal vessel image segmentation; Diffuse-Net includes feature extraction stage and feature fusion stage; In the feature extraction stage, a hybrid differential convolution module HyDC is proposed to comprehensively capture various directional features and edge information of retinal blood vessels and enhance the contrast between blood vessels and background. In the feature fusion stage, an integrated feature fusion module IFFM is designed to integrate high-level features and low-level features of different scales on the basis of grouping aggregation to achieve robust feature representation.

2. The retinal blood vessel segmentation method based on difference feature fusion according to claim 2, characterized in that: The encoder-decoder structure adopted by Diffuse-Net.

3. The retinal blood vessel segmentation method based on difference feature fusion according to claim 1, characterized in that: The hybrid differential convolution module HyDC consists of parallel ordinary convolutions and four differential convolutions. It uses four different types of differential convolutions to enhance gradient-level information and introduces a regular convolution to capture intensity-level information. After obtaining these five different convolution weights, the prior information is provided to the subsequent regular convolution layer.

4. The retinal blood vessel segmentation method based on difference feature fusion according to claim 3, characterized in that: The four differential convolutions include center differential convolution CDC, angle differential convolution ADC, horizontal differential convolution HDC, and vertical differential convolution VDC.

5. The retinal blood vessel segmentation method based on difference feature fusion according to claim 4, characterized in that: The central difference convolution CDC is specifically implemented as follows: In the central difference convolution, a standard 3×3 convolution kernel is first used to obtain normal weights Then, the dimensions of the obtained standard convolution weights are rearranged; after that, a new weight tensor W is created for the central difference convolution cd , and copy the normal weight W to W cd middle; For the non-central position of the center difference convolution kernel, set: W(c in ,c out ,k1,k2)→W(c in ,c out ,k1×k2) W cd (c in ,c out ,k1×k2)=W(c in ,c out ,k1×k2) Among them, c in is the number of input channels, c out is the number of output channels, k1 and k2 are the height and width of the convolution kernel respectively; For the center position of the central difference convolution kernel, its weight is adjusted by subtracting the sum of the weights of all other positions; this step realizes the key transformation from standard convolution to central difference convolution: Among them, i is an index variable used to indicate the position of each element in the 3×3 convolution kernel; Finally, the resized weights are rearranged back to the dimensions of the original convolutional weights.

6. The retinal blood vessel segmentation method based on difference feature fusion according to claim 5, characterized in that: The angle differential convolution ADC is specifically implemented as follows: In the angular differential convolution, a 3×3 convolution kernel is first used to obtain the initial weights and rearrange its dimensions. An additional parameter θ is set to control the strength of the difference so that each position of the convolution weight is adjusted according to θ. The following formula is used to express the process of obtaining the angular differential convolution weights: W ad =(c in ,c out ,i)=W(c in ,c out ,i)-θ·W(c in ,c out ,j(i)) Here, θ is a parameter that controls the differential intensity and j(i) is a function that defines the angular offset with respect to position: mod refers to the remainder operation; for each position i, the weight is adjusted according to the corresponding j(i).

7. The retinal blood vessel segmentation method based on difference feature fusion according to claim 5, characterized in that: The horizontal differential convolution HDC and the vertical differential convolution VDC are specifically implemented as follows: In the horizontal differential convolution, we first use a 3×3 convolution kernel to obtain the standard weights, and create a new tensor of the same shape to store the weights of the horizontal differential convolution. We obtain W by equivalently rearranging the standard weights. hd Specifically, the value of the first column of the standard convolution kernel weight is assigned to the first column of the horizontal difference convolution weight, and the negative value of the third column of the standard convolution kernel weight is assigned to the third column of the horizontal difference convolution weight, and the weight of the horizontal difference convolution kernel is derived. hd : The weights of the vertical difference convolution kernel can be derived similarly by assigning the positive values ​​of the first row and the negative values ​​of the third row of the standard convolution kernel weights to the first and third rows of the vertical difference convolution, respectively:

8. The retinal blood vessel segmentation method based on difference feature fusion according to claim 1, characterized in that: The specific implementation of the integrated feature fusion module IFFM is as follows: The two inputs of the integrated feature fusion module IFFM are: low-level features X l and high-level features X h ; First, use 1×1 convolution and bilinear interpolation to adjust the high-level features X h The number of channels and size are adjusted to obtain the feature X′ h , so that it is consistent with the low-level feature X l With the same number of channels and spatial dimensions: X′ h =Bilinear(Conv 1×1 (X h )) Among them, Bilinear represents the bilinear interpolation operation, Conv 1×1 Represents a 1×1 convolution operation; Then, X′ is transformed along the channel dimension h and X l are divided into four sub-tensors; specifically, X′ h is divided into four sub-tensors: x′ h [0], x′ h [1], x′ h [2] and x′ h [3], X l is divided into four sub-tensors: x l [0], x l [1] x l [2] and x l [3]: x′ h [0],x′ h [1],x′ h [2],x′ h [3]=Split(X′ h ,4,dim=1) x l [0],x l [1],x l [2],xl[3]=Split(x l ,4,dim=1) Among them, Split refers to the chunk function in the Pytorch library; the dim parameter specifies the dimension of the operation, and dim=1 means operating along the channel dimension; Subsequently, the four sub-tensors obtained from the low-level features and high-level features are matched and concatenated along the channel dimension; this process generates four sets of fused features, each containing information from both low-level and high-level features: x′ k =concat(x′ h [k],x l [k]),k∈0,1,2,3 Among them, concat(·) represents the concatenation operation along the channel dimension, i.e., the first dimension; Next, layer normalization is performed on these four sets of fused features, and dilated convolutions with 3×3 convolution kernels and different dilation rates are applied to extract information at different scales, resulting in four sets of information: x0, x1, x2, and x3: x k =ConV d(3×3) (LN(x′ i ))i∈0,1,2,3 Here, d represents the dilation rate, LN refers to layer normalization, ConV d(3×3) Represents a 3×3 convolution operation; Finally, the four sets of information obtained are concatenated, and then layer normalization and 1×1 standard convolution are performed to obtain the final feature X: X=Conv 1×1 (LN(concat(x0,x1,x2,x3)))。 9. The retinal blood vessel segmentation method based on difference feature fusion according to claim 1, characterized in that: This method uses the binary cross entropy-Dice loss function BCEDiceLoss to optimize the network; BCEDiceLoss is a composite loss function that combines the binary cross entropy loss BCE with the Dice loss; Binary cross entropy loss is used to measure the error between the predicted probability and the true label of each pixel for binary classification problems; its formula is: Where N is the total number of pixels, y i is the true label of pixel i, which is 0 or 1, p i is the predicted probability of the i-th pixel; Dice loss measures the similarity between the prediction and the true label by directly optimizing the Dice coefficient, targeting the problem of class imbalance, especially small area foreground segmentation; its formula is: BCEDiceLoss is a combined loss function that is a weighted sum of BCE loss and Dice loss. Its formula is: L BCEDice =α·L BCE +β·L Dice Among them, α and β are adjustable weight parameters used to balance the contribution of BCE loss and Dice loss in the total loss.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 9 can be implemented.

Images