A downsampling method and system for retinal vessel segmentation

Abstract

The present invention discloses a downsampling method for retinal blood vessel segmentation, comprising: first building a basic segmentation framework U-Net network, then replacing the downsampling layer in the original U-Net network model, changing the original Maxpooling layer to a PF-Pooling module, then preprocessing the retinal fundus image to be processed, obtaining a preprocessing result image to be segmented and sending it to the network model, the model extracts pixels at four positions in the sampling window respectively, and then splices the pixels at corresponding positions to obtain four new images with half the scale, then uses Concat stacking to stack the four images in the channel dimension, and finally uses 1*1 convolution to perform channel compression, ultimately achieving the effect of keeping the number of channels of the original image unchanged and halving the width and height, and using this as the downsampling module of the model. Compared with the Maxpooling method, the PF-Pooling sampling method can retain more original image information and further improve the model segmentation accuracy.

Description

Technical Field

[0001] The present invention relates to an image segmentation method, in particular to a downsampling method and system applied to retinal blood vessel segmentation, belonging to the field of image processing. Background Art

[0002] The retinal blood vessels in the fundus are the only vascular images that can be directly and non-invasively observed in the human body. Their morphology and physiological characteristics can reflect the severity of certain diseases to a certain extent. Professional ophthalmologists generally make diagnoses based on this. However, traditional manual screening is too cumbersome and relies heavily on the doctor's subjective consciousness. Therefore, the use of computer-assisted retinal blood vessel segmentation algorithms has gradually become a research focus for fundus screening and diagnosis.

[0003] A recent study demonstrates that deep learning-based methods have achieved excellent results in various areas of medical imaging, including lung segmentation, diabetes identification, brain tumor detection, and retinal vessel segmentation. Currently, numerous algorithmic models have been proposed by domestic and international expert teams for retinal fundus image segmentation. The most classic is the U-Net network model proposed by Ronneberger et al. in 2015. This model uses a symmetrical U-shaped structure as its framework. It extracts features by performing convolution and downsampling operations on the image, and then performs convolution, stacking, and upsampling operations to obtain the segmented image. It has achieved excellent results in various medical image segmentation tasks, and many algorithms are currently improving upon it. Among the most classic are the DenseUnet and ResUnet network models. In 2016, Huang et al. proposed the DenseNet network, which connects all layers to maximize image information extraction while ensuring maximum information transfer between layers. Cai et al. proposed the DenseUnet model by replacing the convolutional modules of the original U-Net network with the DenseBlock module from the DenseNet model. In 2015, He et al. proposed the ResNet network, which effectively solved the gradient explosion and gradient vanishing problems by establishing a correlation channel between the input and output. Diakogiannis et al. [6] proposed the ResUnet model by replacing the convolutional layer of the original U-Net network with the residual module in the ResNet network. Both models increase the width of the network so that the model can extract more feature information to achieve improved segmentation performance. However, while improving performance, they also increase a large number of trainable parameters, which greatly increases the complexity of the model.

[0004] The traditional Maxpooling downsampling method may cause the loss of some image information because it simply selects the local maximum value as the sampling result. The current segmentation method of the existing technology is not ideal. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a downsampling method and system for retinal blood vessel segmentation, which can achieve the effect of retinal blood vessel image segmentation with the number of original image channels unchanged and the width and height halved.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A downsampling method for retinal vessel segmentation includes the following steps:

[0008] Step 1: Build the basic segmentation framework U-Net network, replace the downsampling layer in the original U-Net network model, and feed the image to be segmented into the network model;

[0009] Step 2: Perform convolution operation on the segmented image to extract feature attributes, and then send it to the PF-Pooling module to compress the image;

[0010] Step 3: The PF-Pooling (PixelFusion-Pooling) module performs odd-even splitting, splicing, and channel compression on the original image to obtain an image with the same number of channels but half the width and height;

[0011] Step 4: Continue to perform convolution operation on the image compressed by the PF-Pooling module, and repeat steps 2 and 3 three times to obtain the final feature set;

[0012] Step 5: Upsample the final feature set and concat it with the feature map of the same layer to obtain more original features;

[0013] Step 6: Repeat step 5 three times for the stacked image to obtain an image of the same size as the original image. Finally, compress the channel through 1*1 convolution to obtain the final segmented image.

[0014] As a further improvement of the present invention, the specific steps of step 1 are:

[0015] Step S101: Replace the downsampling layer in the original U-Net network model and change the original Maxpooling layer to a PF-Pooling module. The expression of the PF-Pooling module is as follows:

[0016] y (ij) =K*[x (i,j) ,x (i+1,j) ,x (i,j+1) ,x (i+1,j+1)

[0017] i≤m-2

[0018] j<=n-2

[0019] Where K is the convolution kernel of size 1×1 automatically learned by the model, and m and n are the sampling window sizes;

[0020] Step S102: The image to be segmented is preprocessed before being fed into the network model. A grayscale image is generated using RBG three-channel weighting as the input information of the model, as shown in the following formula:

[0021] Gray=0.299*R+0.587*G+0.114*B#(1)

[0022] Gray is the pixel value of the pixel in the grayscale image, and R, G, and B are the pixel values of the red, green, and blue channels of the original image.

[0023] Step S103: The image to be segmented is pre-processed and then fed into the network model.

[0024] As a further improvement of the present invention, the specific steps of step 2 are:

[0025] Step S201: Perform convolution operation on the image to be segmented;

[0026] Step S202: extracting characteristic attributes of the image to be segmented;

[0027] Step S203: After extracting the feature attributes, the extracted features are sent to the PF-Pooling module to compress the image.

[0028] As a further improvement of the present invention, the specific steps of step three are:

[0029] Step S301: The PF-Pooling module extracts pixels at four positions within the sampling window respectively;

[0030] Step S302: stitching pixels at corresponding positions to obtain four new images with half the scale;

[0031] Step S303: stack the four sub-images in the channel direction to obtain an image with half the scale but four times the depth of the original image;

[0032] Step S304: Use the convolution kernel stack image to perform channel compression to achieve an image with the number of channels of the original image unchanged but the width and height reduced by half.

[0033] As a further improvement of the present invention, the final feature set of step four includes fundus images, images with retinal lesions, and images without retinal lesions, and the final feature set is divided into a training set and a test set.

[0034] As a further improvement of the present invention, the more original features described in step five are specifically as follows: for a neural network, the underlying feature image is often an extraction of texture features of the original image, such as points, lines, edge contours and other information, while the deep feature image is more similar to the semantic information of the original image; when upsampling, the edge features of some feature maps are often lost, and the lost features cannot be retrieved during upsampling, so we concat stack the feature map obtained after upsampling with the downsampled feature map of the same layer, thereby retrieving the edge information of the original image and making it have more original image features.

[0035] As a further improvement of the present invention, the present invention also discloses a downsampling system for retinal vessel segmentation, comprising:

[0036] A module is built to build the basic segmentation framework U-Net network, replace the downsampling layer in the original U-Net network model, and feed the image to be segmented into the network model;

[0037] The extraction module is used to perform convolution operations on the segmented image, extract feature attributes, and then send them to the PF-Pooling module for image compression;

[0038] The compression module is used by the PF-Pooling module to perform odd-even splitting, splicing, and channel compression on the original image, obtaining an image with the same number of channels but half the width and height;

[0039] The convolution module is used to continue the convolution operation on the image compressed by the PF-Pooling module and repeat the steps to obtain the final feature set;

[0040] The stacking module is used to upsample the final feature set and concat it with the feature map of the same layer to obtain more original features;

[0041] The segmentation module is used to compress the channel through 1*1 convolution to obtain the final segmented image.

[0042] As a further improvement of the present invention, the system further includes:

[0043] The preprocessing module is used to preprocess the retinal fundus image to be segmented before it is sent to the network model. The grayscale image is generated by RBG three-channel weighting as the input information of the model to obtain the preprocessing result image.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. Traditional preprocessing generally uses the green channel G of the image as a grayscale image and uses it as the input information of the model. In this invention, the RBG three-channel weighted grayscale image is generated as the input information of the model. The images processed by the preprocessing method of this invention have a more obvious contrast between the blood vessels and the background, making the final segmentation performance of the model more accurate and able to achieve more efficient segmentation of microvessels. Compared with the traditional Maxpooling method, the PF-Pooling method can retain more original image information and further improve the model segmentation accuracy.

[0046] 2. The present invention provides a downsampling method for retinal vessel segmentation. Experimental results on the DRIVE and STARE datasets show that compared with the traditional algorithm model using Max-Pooling, the U-Net model using PF-Pooling improves the F1-score indicator on the STARE dataset by 1.98%, the accuracy by 0.2%, and the sensitivity by 3.88%.

[0047] 3. The generalization of the proposed module was verified by replacing different algorithm models. PF-Pooling achieved performance improvements on both DenseUnet and ResUNet models. PF-Pooling can be well integrated into different segmentation models, and its performance is better than Max-Pooling. The PF-Pooling module has good generalization and can adapt well to different network architectures, thereby improving model performance. It can be effectively applied to medical segmentation tasks and improve segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A diagram showing the steps of the method of the present invention;

[0049] Figure 2 are the four sub-images obtained by performing odd-even splitting on the original image;

[0050] Figure 3 This is the schematic diagram of traditional Max-Pooling sampling;

[0051] Figure 4 This is the PF-Pooling flow chart proposed in the present invention;

[0052] Figure 5 64*64 training data fed into the model ((a) preprocessed image, (b) true blood vessel image);

[0053] Figure 6 Comparison chart of different preprocessing results;

[0054] Figure 7The performance of PF-Pooling on the U-Net model;

[0055] Figure 8 Verification of the generalization of the DSB2018 dataset ((a) original image; (b) true value image;

[0056] (c) U-Net (Max-pooling) segmentation result; (d) U-Net (PF-Pooling) segmentation result). DETAILED DESCRIPTION

[0057] The present invention will be further described in detail below by way of examples. The following examples are intended to explain the present invention but the present invention is not limited to the following examples.

[0058] The embodiment of the present invention provides a downsampling method for retinal blood vessel segmentation, such as Figure 1 As shown, the method includes the following steps:

[0059] Step 1: Build the basic segmentation framework U-Net network, replace the downsampling layer in the original U-Net network model, and feed the image to be segmented into the network model;

[0060] Step 2: Perform convolution operation on the segmented image to extract feature attributes, and then send it to the PF-Pooling module to compress the image;

[0061] Step 3: The PF-Pooling module performs odd-even splitting, splicing, and channel compression on the original image to obtain an image with the same number of channels but half the width and height;

[0062] Step 4: Continue to perform convolution operation on the image compressed by the PF-Pooling module, and repeat steps 2 and 3 three times to obtain the final feature set;

[0063] Step 5: Upsample the final feature set and concat it with the feature map of the same layer to obtain more original features;

[0064] Step 6: Repeat step 5 three times for the stacked image to obtain an image of the same size as the original image. Finally, compress the channel through 1*1 convolution to obtain the final segmented image.

[0065] The method steps of the present invention are as follows: Figure 2As shown in the figure, image is the original image, which is split into rows and columns with odd and even numbers, and the pixels at the corresponding positions are extracted respectively to obtain four images (image0: odd rows and odd columns, image1: odd rows and even columns, image2: even rows and odd columns, image3: even rows and even columns). The information of the obtained sub-images is basically consistent with the original image. Inspired by this, the present invention considers the traditional Max-Pooling (such as Figure 3 As shown in Figure 2, since the method of simply selecting the local maximum value as the sampling result may cause the loss of some image information, the present invention constructs a module that can fuse the adjacent pixel information to achieve downsampling. The model architecture is as follows: Figure 4 As shown. Unlike the Max-Pooling method that only takes the maximum value within a 2*2 window, the model in the present invention extracts pixels at four positions within the sampling window separately, and then splices the pixels at the corresponding positions to obtain four new images with half the scale. The four images with half the scale are then stacked in the channel dimension, and finally 1*1 convolution is used for channel compression. Ultimately, the number of channels of the original image remains unchanged, while the width and height are halved. This is used as the downsampling module of the model. Compared with the Max-Pooling method, the proposed PF-Pooling method can retain more original image information and further improve the model segmentation accuracy.

[0066] In an optional implementation manner, the specific steps of step 1 are:

[0067] Step S101: Replace the downsampling layer in the original U-Net network model and change the original Maxpooling layer to a PF-Pooling module. The expression of the PF-Pooling module is as follows:

[0068] y (ij) =K*[x (i,j) ,x (i+1,j) ,x (i,j+1) ,x (i+1,j+1 )]

[0069] i≤m-2

[0070] j<=n-2#(1)

[0071] Where K is the convolution kernel of size 1×1 automatically learned by the model, and m and n are the sampling window sizes;

[0072] Step S102: The image to be segmented is preprocessed before being fed into the network model. A grayscale image is generated using RBG three-channel weighting as the input information of the model, as shown in the following formula:

[0073] Gray=0.299*R+0.587*G+0.114*B#(1)

[0074] Gray is the pixel value of the pixel in the grayscale image, and R, G, and B are the pixel values of the red, green, and blue channels of the original image.

[0075] Step S103: The image to be segmented is pre-processed and then fed into the network model.

[0076] Among them, the present invention first builds a traditional U-Net segmentation model, and then replaces the downsampling module in the model with the PF-Pooling module proposed in the present invention (Formula 1), where K is the convolution kernel of size 1×1 automatically learned by the model, and m and n are the sampling window sizes.

[0077] y (ij) =K*[x (i,j) ,x (i+1,j) ,x (i,j+1) ,x (i+1,j+1 )]

[0078] i≤m-2

[0079] j<=n-2

[0080] Compared with the traditional Max-Pooling module (Formula 2), this module can better fuse adjacent pixel information, retain more original image information, and improve the model segmentation accuracy.

[0081] Since retinal fundus images often have uneven brightness information, the contrast between small blood vessels and the background is not obvious, which is not conducive to model segmentation. In order to better realize the extraction of small blood vessels, the original image needs to be preprocessed first. Traditional preprocessing generally uses the green channel G of the image as a grayscale image and uses it as the input information of the model. The present invention uses RBG three-channel weighted grayscale image (as shown in formula (3)) as the input information of the model, where Gray is the pixel value of the pixel in the grayscale image, and R, G, and B are the pixel values of the red, green, and blue three-channel pixels of the original image. After that, the contrast between blood vessels and background information is further improved through image normalization, histogram equalization, and gamma correction.

[0082] Gray=0.299*R+0.587*G+0.114*B#(1)

[0083] In an optional implementation manner, the specific steps of step 2 are:

[0084] Step S201: Perform convolution operation on the image to be segmented;

[0085] Step S202: extracting characteristic attributes of the image to be segmented;

[0086] Step S203: After extracting the feature attributes, the extracted features are sent to the PF-Pooling module to compress the image.

[0087] Among them, the convolution operation is performed on the image to be segmented, and a 1*1 convolution kernel is used for convolution operation; then the retinal vascular feature attributes of the image to be segmented are extracted; finally, the extracted retinal vascular feature attributes are sent to the PF-Pooling module to compress the image.

[0088] In an optional implementation manner, the specific steps of step three are:

[0089] Step S301: The PF-Pooling module extracts pixels at four positions within the sampling window respectively;

[0090] Step S302: stitching pixels at corresponding positions to obtain four new images with half the scale;

[0091] Step S303: stack the four sub-images in the channel direction to obtain an image with half the scale but four times the depth of the original image;

[0092] Step S304: Use the convolution kernel stack image to perform channel compression to achieve an image with the number of channels of the original image unchanged but the width and height reduced by half.

[0093] The present invention constructs a module that can fuse adjacent pixel information to achieve downsampling. The model architecture is as follows Figure 4 Unlike the Max-Pooling method, which only takes the maximum value within a 2*2 window, the model in this paper extracts pixels at four positions within the sampling window separately, then concatenates the pixels at corresponding positions to obtain four new images with half the scale. These four images with half the scale are then stacked in the channel dimension, and finally 1*1 convolution is used for channel compression, ultimately achieving the effect of halving the width and height of the original image while keeping the number of channels unchanged.

[0094] In an optional implementation manner, the final feature set of step four includes fundus images, images with retinal lesions, and images without retinal lesions, and the final feature set is divided into a training set and a test set.

[0095] Among them, such as Figure 5 As shown, the fundus images, images with retinal lesions, and images without retinal lesions of the present invention are as follows using the DRIVE dataset and the STARE dataset:

[0096] DRIVE:

[0097] The images in this dataset come from a diabetic retinopathy screening program in the Netherlands, which screened 400 people with diabetes aged 25-90. Forty of these images were randomly selected to form the DRIVE dataset, which includes 33 images without diabetic retinopathy and 7 images with mild early-stage retinopathy. Twenty of these images were used as the training set, and 20 as the test set.

[0098] STARE:

[0099] The data set contains 20 fundus images, 10 of which show retinal lesions and 10 do not. 15 of them are used as the training set and 5 as the test set.

[0100] The real blood vessel segmentation maps for both datasets were manually segmented by experienced ophthalmologists. To further expand the dataset, the pre-processed images were rotated and inverted, and 200 64*64 regions of interest (ROI) were randomly selected from the circular region of interest (ROI) within each image and fed into the model for training.

[0101] An optional implementation method, the more original features described in step five are specifically as follows: for a neural network, the underlying feature image is often an extraction of texture features of the original image, such as points, lines, edge contours and other information, while the deep feature image is more similar to the semantic information of the original image; when upsampling, the edge features of some feature maps are often lost, and the lost features cannot be retrieved during upsampling, so we concat stack the feature map obtained after upsampling with the downsampled feature map of the same layer, thereby retrieving the edge information of the original image and making it have more original image features.

[0102] In an optional embodiment, the system includes:

[0103] The preprocessing module is used to preprocess the retinal fundus image to be segmented before it is sent to the network model. The grayscale image is generated by RBG three-channel weighting as the input information of the model to obtain the preprocessing result image.

[0104] The model uses a weighted grayscale image generated by the RBG three-channel method (as shown in formula (1)), where Gray represents the pixel value of the grayscale image, and R, G, and B represent the pixel values of the red, green, and blue channels of the original image. Image normalization, histogram equalization, and gamma correction are then performed to further enhance the contrast between blood vessels and background information.

[0105] Gray=0.299*R+0.587*G+0.114*B#(1)

[0106] By comparing the input image as an RGB three-channel image, the image after grayscale conversion using the G channel, and the image after weighted grayscale conversion using the three channels used in the present invention, as shown in FIG. Figure 6 It can be seen from the various indicators shown that the contrast between the blood vessels and the background of the images processed by the preprocessing method of the present invention is more obvious, which makes the final segmentation performance of the model more accurate and can achieve more efficient segmentation of microvessels.

[0107] An optional implementation method includes building a module for building a basic segmentation framework U-Net network, replacing the downsampling layer in the original U-Net network model, and feeding the image to be segmented into the network model;

[0108] The extraction module is used to perform convolution operations on the segmented image, extract feature attributes, and then send them to the PF-Pooling module for image compression;

[0109] The compression module is used by the PF-Pooling module to perform odd-even splitting, splicing, and channel compression on the original image, obtaining an image with the same number of channels but half the width and height;

[0110] The convolution module is used to continue the convolution operation on the image compressed by the PF-Pooling module and repeat the steps to obtain the final feature set;

[0111] The stacking module is used to upsample the final feature set and concat it with the feature map of the same layer to obtain more original features;

[0112] The segmentation module is used to compress the channel through 1*1 convolution to obtain the final segmented image.

[0113] The system module described in the present invention evaluates the proposed PF-Pooling module on different models. First, an original U-Net network is built, and the Max-Pooling module is then replaced with the PF-Pooling module proposed in the present invention. The final segmented image is then passed through the preprocessing module, compression module, processing module, convolution module, stacking module, and segmentation module. The models using the PF-Pooling method and the Max-Pooling method are then compared on different datasets and models, further demonstrating the effectiveness and universality of the PF-Pooling module.

[0114] The evaluation criteria are as follows:

[0115] In order to accurately compare the performance of different models, four different evaluation indicators are used for comparison, namely the area under the ROC curve AUC, accuracy Acc (as shown in formula (2)), F1-Score (as shown in formula (3)) and sensitivity Sen (as shown in formula (4))

[0116]

[0117] Where TP, TN, FP, and FN represent the number of times a point is predicted as a vessel and correctly predicted, the number of times a point is predicted as background and correctly predicted, the number of times a point is predicted as a vessel and incorrectly predicted, and the number of times a point is predicted as background and incorrectly predicted, respectively. precision represents the precision rate (as shown in formula (5)), and recall represents the recall rate (as shown in formula (6)).

[0118]

[0119] According to the evaluation criteria, the comparison results are shown in Table 1 (the black part is the best result within the indicator). It can be seen from the table that the proposed PF-Pooling is superior to the U-Net model using traditional Max-Pooling in all indicators, and obtains a huge improvement of 1.93% in the F1-score indicator of the STARE dataset, which means that PF-Pooling can more effectively improve the segmentation accuracy of microvessels and Figure 7 It can be seen that the network model using PF-Pooling performs better than the network model using the original Max-Pooling in each round of iteration and converges faster.

[0120] Table 1 Comparison of various indicators of the model

[0121]

[0122] The present invention also verifies the generalization of the PF-Pooling module on other newly proposed models. The results are shown in Table 2 (the bold parts are the optimal results within the indicators). From the various indicators in the table, it can be seen that the network model using PF-Pooling as the downsampling module has improved in all indicators. This shows that the PF-Pooling module proposed in the present invention has good generalization and can adapt well to different network architectures, thereby improving model performance.

[0123] Table 2 PF-Pooling generalization verification

[0124]

[0125] Finally, in order to further verify the generalization of the PF-Pooling module, the present invention uses the U-Net network model for evaluation on the DSB2018 dataset. Figure 8 From the upper left white circle shown in (c) and (d), we can see that the U-Net network using the PF-Pooling module can better fuse the semantic information of adjacent pixels to achieve accurate segmentation, which further proves that the PF-Pooling module has good generalization performance.

[0126] This paper proposes a downsampling module PF-Pooling with good generalization. Compared with the traditional Max-Pooling module, it can better fuse the semantic information between adjacent pixels. The proposed model is verified in retinal vessel segmentation tasks and cell segmentation tasks.

[0127] Experimental results show that the model using PF-Pooling as the downsampling module outperforms the network model using the Max-Pooling module in segmentation performance, and can more effectively extract small blood vessels in the fundus. Experimental results on different models and different segmentation tasks also show that the proposed PF-Pooling module has good generalization properties and can be effectively applied to medical segmentation tasks to improve segmentation accuracy.

[0128] The above contents described in this specification are merely examples of the present invention. Those skilled in the art may make various modifications, additions, or substitutions to the described embodiments, without departing from the contents of this specification or exceeding the scope defined by the claims, and such modifications, additions, or substitutions may be made to the described embodiments. Such modifications, additions, or substitutions may be made by persons skilled in the art. Such modifications, additions, or substitutions may be made to the described embodiments without departing from the contents of this specification or exceeding the scope defined by the claims, and such modifications shall fall within the scope of protection of the present invention.

Claims

1. A downsampling method for retinal vessel segmentation, characterized in that: The following steps are involved: Step 1: Build the basic segmentation framework U-Net network, replace the downsampling layer in the original U-Net network model, and feed the image to be segmented into the network model; Step 2: Perform convolution operation on the segmented image to extract feature attributes, and then send it to the PF-Pooling module to compress the image; Step 3: The PF-Pooling module performs odd-even splitting, splicing, and channel compression on the original image to obtain an image with the same number of channels but half the width and height; Step 4: Continue to perform convolution operation on the image compressed by the PF-Pooling module, and repeat steps 2 and 3 three times to obtain the final feature set; Step 5: Upsample the final feature set and concat it with the feature map of the same layer to obtain more original features; Step 6: Repeat step 5 three times for the stacked image to obtain an image of the same size as the original image. Finally, compress the channel through 1*1 convolution to obtain the final segmented image. The specific steps of step 1 are: Step S101: Replace the downsampling layer in the original U-Net network model and change the original Maxpooling layer to a PF-Pooling module. The expression of the PF-Pooling module is as follows: y (ij) =K*[x (i,j) ,x (i+1,j) ,x (i,j+1) ,x (i+1,j+1) ] i≤m-2 j<=n-2 Where K is the convolution kernel of size 1×1 automatically learned by the model, and m and n are the sampling window sizes; Step S102: The image to be segmented is preprocessed before being fed into the network model. A grayscale image is generated using RBG three-channel weighting as the input information of the model, as shown in the following formula: Gray=0.299*R+0.587*G+0.114*B#(1) Gray is the pixel value of the pixel in the grayscale image, and R, G, and B are the pixel values of the red, green, and blue channels of the original image. Step S103: The image to be segmented is pre-processed and then fed into the network model; The specific steps of step three are: Step S301: The PF-Pooling module extracts pixels at four positions within the sampling window respectively; Step S302: stitching pixels at corresponding positions to obtain four new images with half the scale; Step S303: stack the four sub-images in the channel direction to obtain an image with half the scale but four times the depth of the original image; Step S304: Use a convolution kernel to perform channel compression on the image to achieve an image with the number of channels of the original image unchanged but the width and height reduced by half.

2. The downsampling method for retinal vessel segmentation according to claim 1, characterized in that: The specific steps of step 2 are: Step S201: Perform convolution operation on the image to be segmented; Step S202: extracting characteristic attributes of the image to be segmented; Step S203: After extracting the feature attributes, the extracted features are sent to the PF-Pooling module to compress the image.

3. The downsampling method for retinal vessel segmentation according to claim 1, characterized in that: The final feature set of step 4 includes fundus images, images with retinal lesions, and images without retinal lesions. The final feature set is divided into a training set and a test set.

4. The downsampling method for retinal vessel segmentation according to claim 1, characterized in that: The more original features described in step five are specifically as follows: for a neural network, the feature image of the bottom layer is often an extraction of the texture features of the original image, while the feature image of the deep layer is closer to the semantic information of the original image; when upsampling, the edge features of some feature maps are often lost, and the lost features cannot be retrieved during upsampling, so we concat stack the feature map obtained after upsampling with the downsampled feature map of the same layer, thereby retrieving the edge information of the original image and making it have more original image features.

5. A system for the downsampling method for retinal vessel segmentation according to claim 1, characterized in that: A module is built to build the basic segmentation framework U-Net network, replace the downsampling layer in the original U-Net network model, and feed the image to be segmented into the network model; The extraction module is used to perform convolution operations on the segmented image, extract feature attributes, and then send them to the PF-Pooling module for image compression; The compression module is used by the PF-Pooling module to perform odd-even splitting, splicing, and channel compression on the original image, obtaining an image with the same number of channels but half the width and height; The convolution module is used to continue the convolution operation on the image compressed by the PF-Pooling module and repeat the steps to obtain the final feature set; The stacking module is used to upsample the final feature set and concat it with the feature map of the same layer to obtain more original features; The segmentation module is used to compress the channel through 1*1 convolution to obtain the final segmented image.

6. The downsampling system for retinal vessel segmentation according to claim 5, further comprising: The preprocessing module is used to preprocess the retinal fundus image to be segmented before it is sent to the network model. The grayscale image is generated by RBG three-channel weighting as the input information of the model to obtain the preprocessing result image.

Images