CN106920227A - Based on the Segmentation Method of Retinal Blood Vessels that deep learning is combined with conventional method - Google Patents

Abstract

A retinal vessel segmentation method based on the combination of deep learning and traditional methods, involving the fields of computer vision and pattern recognition. The present invention uses two kinds of grayscale images as training samples of the network, and performs corresponding data amplification including elastic deformation, smoothing filtering, etc. for the problem of less retinal image data, which expands the wide applicability of the present invention. In the present invention, by constructing the retinal vessel segmentation depth network of FCN‑HNED, the network realizes the process of autonomous learning to a great extent, not only can share the convolution features of the entire image, reduce feature redundancy, but also recover from abstract features The classification of multiple pixels is obtained, and the CLAHE map and Gaussian matching filter map of the retinal blood vessel image are respectively input into the network so that the obtained blood vessel segmentation map is weighted and averaged to obtain a better and more complete retinal blood vessel segmentation probability map. This method greatly improves the robustness and accuracy of vessel segmentation.

Description

Retinal Vessel Segmentation Method Based on Combination of Deep Learning and Traditional Methods

technical field

The invention relates to the fields of computer vision and pattern recognition, and is a retinal blood vessel segmentation method based on the combination of deep learning and traditional methods.

Background technique

Fundus imaging can be used to determine whether there is an abnormality through retinal imaging, and the observation of retinal blood vessels is very important. Diseases such as glaucoma, cataract, and diabetes can all cause lesions in the blood vessels of the retinal fundus. The number of patients with retinopathy is increasing year by year, and if it is not treated in time, it usually leads to great pain and even blindness in patients with long-term suffering from these diseases. However, at present, retinopathy is manually diagnosed by a specialist. The specialist first manually marks the blood vessels on the patient's fundus image, and then measures the required blood vessel diameter, bifurcation angle and other related parameters. Among them, the process of manually marking blood vessels takes about two hours, and the diagnosis process takes a lot of time. In order to save manpower and material resources, the method of automatically extracting blood vessels is particularly important. Not only can it reduce the burden on specialists, but it can also effectively solve the problem of lack of specialists in remote areas. In view of the importance of retinal vessel segmentation, domestic and foreign scholars have done a lot of research, roughly divided into unsupervised and supervised methods.

The unsupervised method is to extract blood vessel targets through certain rules, including algorithms such as matched filtering, morphological processing, blood vessel tracking, and multi-scale analysis. Supervised learning is also called pixel feature classification method or machine learning technique. Through training, each pixel is classified as vascular or non-vascular. It is mainly divided into two processes: feature extraction and classification. The feature extraction stage usually includes methods such as Gabor filtering, Gaussian matching filtering, and morphological enhancement. Classifiers usually included in the classification stage include Bayesian (naive Bayesian), SVM and other classifiers. However, this kind of judgment for pixels cannot well consider the connection between each pixel and its surrounding domain pixels. Therefore, CNN appeared, which can judge whether the central pixel is a blood vessel or non-vascular according to the characteristics of the image block, and through the automatic learning of features of the multi-layer structure, these abstract features are conducive to the classification and judgment of the central pixel. However, classifying each pixel rarely involves global information, so that the classification fails in the case of local lesions; secondly, each image has at least hundreds of thousands of pixels, and if it is judged one by one, the storage overhead will be large, and the calculation low productivity.

Contents of the invention

Aiming at the shortcomings of existing algorithms, the present invention proposes a retinal vessel segmentation method based on the combination of deep learning and traditional methods. First of all, according to the characteristics of retinal blood vessels, do targeted preprocessing, including CLAHE (Constrained Contrast Adaptive Histogram Equalization) processing so that retinal blood vessels and background can have a high contrast, and the traditional method of Gaussian matched filtering makes the retina The small blood vessels are well enhanced, and the present invention proposes to use both grayscale images as training samples for the network. On this basis, we have made corresponding data amplification including elastic deformation, smoothing filtering, etc. for the problem of less retinal image data, which not only makes the increase in data volume conducive to the learning and training of deep learning networks, but more importantly, simulates Even if there are retinal images in various situations, a good retinal blood vessel segmentation map can be obtained through the processing of the present invention, which expands the wide applicability of the present invention.

Secondly, by constructing the retinal vessel segmentation depth network of FCN-HNED, the present invention combines the vessel probability map obtained at the end of the FCN (Fully Convolutional Network) network with the vessel probability map of shallow information HNED (HolisticallyNestedEdge Detection), and obtains our The required retinal vessel segmentation map, the network greatly realizes the process of autonomous learning, not only can share the convolution features of the entire image, reduce feature redundancy, but also recover the ownership of multiple pixels from abstract features category, implementing an end-to-end, pixel-to-pixel retinal vessel segmentation method, which is both simple and effective with global input and global output. In the detection of retinal blood vessels, the present invention respectively inputs the CLAHE map and Gaussian matching filter map of the retinal blood vessel image into the network so that the obtained blood vessel segmentation map is weighted and averaged to obtain a better and more complete retinal blood vessel segmentation probability map. This processing method The robustness and accuracy of vessel segmentation are greatly improved.

This article adopts the following technical solutions:

1. Pretreatment

1) Extract the green channel with relatively high contrast in the RGB three channels of the color retinal image. Secondly, due to problems such as shooting angles, the brightness of the collected retinal fundus images is often uneven, or the lesion area is difficult to distinguish from the background due to problems such as too bright or too dark and low contrast in the image. Therefore, we Perform normalization. Then, CLAHE processing is performed on the normalized retinal image to improve the quality of the retinal fundus image, balance the brightness of the fundus image, and make it more suitable for subsequent blood vessel extraction.

The retinal blood vessels processed by CLAHE can maintain the characteristics of the retinal blood vessels to a great extent while enhancing the contrast between the blood vessels and the background. However, because the small blood vessels are very similar to the background, they cannot be well segmented in the subsequent deep learning. In view of this, the present invention uses the Gaussian characteristic of the cross-sectional grayscale image of blood vessels, and performs Gaussian matching filter processing on the retinal blood vessels after CLAHE processing, so that the small blood vessels can be displayed to a great extent. Since the direction of blood vessels is arbitrary, this paper uses 12 Gaussian kernel templates in different directions to perform matching filtering on retinal images, and takes the maximum response as the response value of the pixel. The two-dimensional Gaussian matched filter kernel function k(x,y) can be expressed as:

Where σ represents the variance of the Gaussian curve, L represents the length of the truncated retinal blood vessels on the y-axis, the width of the filter window is [-3σ, 3σ], which is the value range of the kernel function x, and the smaller value of σ is set to 0.5, so that Small blood vessels can be greatly enhanced.

In order to fully consider the overall characteristics of the retinal image and the characteristics of the small blood vessels in it, we use the retinal blood vessel map after CLAHE processing and the Gaussian matched filter map as training samples, which can greatly improve the performance of network segmentation.

2. Data amplification and construction of training samples

Since training a deep network requires a large amount of data, only existing retinal images are not enough for training. Therefore, it is necessary to expand the training data in different ways, increase the amount of data, and improve the training and detection effects. Data augmentation method:

1) Translate the preprocessed image by 20 pixels from left to right, up and down, respectively, to realize the translation invariance of network learning.

2) Rotate the processed image by 45°, 90°, 125°, and 180° respectively, and intercept the largest rectangle among them. This transformation not only enhances the rotation robustness of the training data, but also expands the data to 5 times the original.

3) In the general data amplification, the possible blurring of the retinal image has never been considered. However, the present invention considers that under various circumstances, such as camera shaking or patient’s careless movement, the retinal image will be blurred in a certain amount. degree of partial blurring, so the present invention selects 25% of the processed image sets to perform 3×3 and 5×5 median filter fuzzy operations respectively, so that the network can have retinal images with various blurring degrees. Broad applicability.

4) In the previous retinal image data amplification, only translation, zooming, rotation, etc. are commonly used, which are far from the consideration of various conditions of the retinal image. In view of this, the present invention considers the shape of the blood vessel direction of the retina, etc. Anisotropy, we take 25% of the processed image set for 3) and perform random elastic deformation. This data amplification method is of great significance for the segmentation of retinal blood vessels. It can help the network to learn the intricate retina in various directions. Blood vessels are beneficial to improve the accuracy of retinal blood vessel segmentation in practical applications.

5) Since FCN is applicable to images of any size, we perform 50% and 75% scaling on the processed images of 4) to amplify the data.

Of course, we do the same for the expert standard map (ground truth) of retinal vessel segmentation, so as to correspond to the samples one by one. 3/4 of the well-built training sample data is used as the training set, and 1/4 is used as the verification set.

3. FCN-HNED network construction

FCN network: The general FCN network layer is mainly composed of 5 parts, the input layer, the convolution layer, the downsampling layer, the upsampling layer (deconvolution layer) and the output layer. The network constructed in the present invention is:

Input layer, two convolutional layers (C1, C2), first downsampling layer (pool1), two convolutional layers (C3, C4), second downsampling layer (pool2), two convolutional layers (C5 , C6), the third downsampling layer (pool3), two convolutional layers (C7, C8), the fourth downsampling layer (pool4), two convolutional layers (C9, C10), the first upsampling layer ( U1), two convolutional layers (C11, C12), second upsampling layer (U2), two convolutional layers (C13, C14), third upsampling layer (U3), two convolutional layers (C15 , C16), the fourth upsampling layer (U4), two convolutional layers (C17, C18), and the target layer (output layer). Form a symmetrical U-shaped deep network architecture.

Since the low-level feature resolution of the FCN network is high, and the high-level information reflects stronger semantic information, it is very robust for the classification of blood vessels in some lesions of the retinal image, but at the same time, the FCN network finally gets the same as the input The output of the same sample size will lose a lot of small target and local detail information. Therefore, the present invention uses the edge detection HNED (Holistically nested edge detection) method to learn rich multi-dimensional information under deep supervision. Layer information expression, to a large extent, solves the problem of blurring the edge of the target. That is, we add a softmax classifier after the C2, C4, C6, and C8 layers, so that the information of the hidden layer can be learned to obtain the retinal blood vessel probability map when the groundtruth is used as the label, which are called side output 1 and side output 2 respectively. , Side output 3, Side output 4. On this basis, we fuse the four side outputs with the final output layer to form the network structure of FCN-HNED, complement the shallow layer information with the output layer information, and obtain multi-scale, multi-level, and target samples. The similar fusion feature maps play a great role in the refinement of segmented vessels, so that no subsequent special refinement steps are required for the refinement of retinal vessels.

The convolutional layers of the present invention obtain feature maps of the same size by padding zeros, and the result of the pooling layer is to reduce features and parameters, but the purpose of the pooling layer is not limited to this. The present invention uses max-pooling to reduce the deviation of the estimated mean value caused by the parameter error of the convolution layer, and retain more texture information. The sampling rate of the maximum pooling layer in the present invention is 2. Upsampling is the process of bilinear interpolation.

During the construction of the entire model, the activation function uses ReLU to remove the Softmax classification layer, and the loss function is cross entropy.

Training: After the FCN-HNED network is constructed, the network can be trained to perform automatic feature extraction and learning of the image. Each generation inputs 128 images until the network converges and stops.

Test: Input the CLAHE map and Gaussian matching filter map of each retinal image green channel map to the trained network for testing, and obtain the fused retinal vessel segmentation map called with right with A weighted average is performed to obtain the final retinal vessel segmentation probability map. 4. Post-processing

Binarize the retinal blood vessel probability map obtained in the test to obtain a segmentation map.

Beneficial effect

1. According to the different characteristics of retinal blood vessels, the present invention adopts a targeted data processing method. The quality of the training data directly determines whether the model obtained through training is reliable and whether the accuracy reaches the required level. The present invention uses fuzzy operation, elastic Deformation, etc., well simulates a variety of possible retinal data, and at the same time expands the data to a sufficient number to avoid training overfitting, and can also provide assistance for subsequent detection, thereby improving the accuracy of retinal blood vessel segmentation.

2. In the present invention, the retinal image processed by CLAHE and the image processed by Gaussian matching filter are respectively input into the network for training and learning, which not only enables the properties of retinal blood vessels to be fully learned at each performance level, but also the Gaussian matched filter image fully compensates for the CLAHE The lack of clearness of small blood vessels in the processing map greatly improves the performance of retinal blood vessel segmentation.

3. The present invention can quickly perform automatic feature extraction of retinal images through the method of constructing a deep learning network FCN-HNED. It can perform feature extraction on retinal fundus images from different levels, and learn each pixel in the retinal image and its surroundings The relationship between multiple neighborhoods shows the medium and high-level features of the retinal blood vessel map well, so that it can distinguish the internal features of blood vessels and non-vessels well, and realize end-to-end, pixel-to-pixel The blood vessel segmentation is many times more efficient than the classification and judgment of traditional single pixels.

4. In the present invention, the four side outputs of shallow features are highly fused with the terminal output of the FCN network, so as to achieve refinement and robustness of blood vessel segmentation. The blood vessel segmentation map is in good agreement with the manual segmentation map of experts. At the same time, the automation of retinal vessel segmentation is greatly realized, which greatly reduces the consumption of manpower and material resources.

Description of drawings

Fig. 1 is the overall flowchart of the present invention;

Figure 2 is the gray distribution map of the blood vessel cross section; (a) a section of the blood vessel (b) the gray level

Figure 3 is a preprocessing effect diagram; (a) original image (b) image after CLAHE processing (c) image after Gaussian matching filter

Figure 4 is the FCN-HNED network structure;

Figure 5 is the result of retinal vessel segmentation. (a) Original image (b) Segmentation map of retinal vessels (c) Manual segmentation map by the first expert

detailed description

Describe in detail below in conjunction with accompanying drawing:

The technical block diagram of the present invention is as shown in Figure 1. The specific implementation steps are as follows:

1. Pretreatment

The same preprocessing is performed on each retinal fundus image whether it is a training set or a test set.

1) Extract the green channel with relatively high contrast in the RGB three channels of the color retinal image. Secondly, due to problems such as shooting angles, the brightness of the collected retinal fundus images is often uneven, or the lesion area is difficult to distinguish from the background due to problems such as too bright or too dark and low contrast in the image. Therefore, we Perform normalization processing, and then perform CLAHE processing on the normalized retinal image to improve the quality of the retinal fundus image, balance the brightness of the fundus image, and make it more suitable for subsequent blood vessel extraction.

The retinal blood vessels processed by CLAHE can maintain the characteristics of the retinal blood vessels to a great extent while enhancing the contrast between the blood vessels and the background. However, because the small blood vessels are very similar to the background, they cannot be well segmented in the subsequent deep learning. In view of this, the present invention uses the Gaussian characteristic of the cross-sectional grayscale image of the blood vessel to perform matching filtering processing on the retinal image. As shown in Figure 2, (a) is the blood vessel grayscale image, (b) is the grayscale value of the cross section of the blood vessel, and the cross section of the small blood vessel also presents a Gaussian trend, so the present invention treats the retinal blood vessel after CLAHE Perform Gaussian matched filtering. Since the direction of blood vessels is arbitrary, this paper uses 12 Gaussian kernel templates in different directions to perform matching filtering on retinal images, and finds the corresponding maximum response as the response value of the pixel.

In order to fully consider the overall characteristics of the retinal image and the characteristics of the small blood vessels in it, we use the retinal blood vessel map after CLAHE processing and the Gaussian matched filter map as training samples, which can greatly improve the performance of network segmentation.

2. Data amplification and construction of training samples

Since training a deep network requires a large amount of data, only existing retinal images are not enough for training. Therefore, it is necessary to expand the training data in different ways, increase the amount of data, and improve the training and detection effects. Data augmentation method:

1) Translate the preprocessed image by 20 pixels from left to right, up and down, respectively, to realize the translation invariance of network learning.

2) Rotate the processed image by 45°, 90°, 125°, and 180° respectively, and intercept the largest rectangle among them. This transformation not only enhances the rotation robustness of the training data, but also expands the data to 5 times the original.

3) In general data augmentation, the median filter has never been used. However, the present invention considers that under various circumstances, such as camera shake or patient’s careless movement, the retinal image will be distorted to a certain extent. Therefore, the present invention takes 25% of the processed images to perform 3×3 and 5×5 median filter fuzzy operations respectively, so that the network can be widely applicable to retinal images with various blurring degrees sex.

4) In the previous retinal image data amplification, only translation, zooming, rotation, etc. are commonly used, which are far from the consideration of various conditions of the retinal image. In view of this, the present invention considers the shape of the blood vessel direction of the retina, etc. Anisotropy, we take 25% of the processed image for 3) and perform random elastic deformation. This data amplification method is of great significance for the segmentation of retinal blood vessels. It can help the network to learn intricate retinal blood vessels in various directions. , which is conducive to the improvement of the accuracy of retinal vessel segmentation in practical applications.

5) Since FCN is applicable to images of any size, we perform 50% and 75% scaling on the processed images of 4) to amplify the data.

Of course, we do the same for the expert standard map (ground truth) of retinal vessel segmentation, so as to correspond to the samples one by one. 3/4 of the well-built training sample data is used as the training set, and 1/4 is used as the verification set.

3. FCN-HNED network construction and training and testing process

FCN network: The general FCN network layer is mainly composed of 5 parts, the input layer, the convolution layer, the downsampling layer, the upsampling layer (deconvolution layer) and the output layer. The network constructed in the present invention is: input layer, two convolutional layers (C1, C2), the first downsampling layer (pool1), two convolutional layers (C3, C4), the second downsampling layer (pool2) , two convolutional layers (C5, C6), third downsampling layer (pool3), two convolutional layers (C7, C8), fourth downsampling layer (pool4), two convolutional layers (C9, C10 ), the first upsampling layer (U1), two convolutional layers (C11, C12), the second upsampling layer (U2), two convolutional layers (C13, C14), the third upsampling layer (U3) , two convolutional layers (C15, C16), fourth upsampling layer (U4), two convolutional layers (C17, C18), target layer (output layer). Form a symmetrical U-shaped deep network architecture.

The convolution process is implemented as follows:

f(X;W,b)=W* _s X+b (2)

Among them, f(X; W, b) is the output feature map, X is the input feature map of the previous layer, W and b are the convolution kernel and offset value, * _s represents the convolution operation, unlike the traditional CNN Network, the FCN network replaces all the last fully connected layers with convolutional layers. However, after a series of operations such as convolution and downsampling, the feature map becomes smaller and smaller. To restore the image to the same size as the input image, FCN uses Upsampling operation or deconvolution.

The intermediate convolutional layers of the present invention obtain feature maps of the same size by padding zeros, and in the left-right symmetrical U-shaped network, two closely connected 3×3 filter convolution kernels are repeatedly applied for convolution operations, with a step size of 1. There is a ReLU activation function behind each convolutional layer. The result of the pooling layer is to reduce features and parameters, but the purpose of the pooling layer is not limited to this. It can maintain some invariance rotation, translation, etc., this The structure uses a max-pooling layer with a core of 2×2 and a step size of 2, which can reduce the deviation of the estimated mean value caused by the parameter error of the convolution layer, and retain more texture information. During each downsampling, the number of feature maps doubles, and upsampling does the opposite. In addition, a 1×1 convolution kernel is used in the last layer to map 64 feature maps to the standard output for training.

During the construction of the entire model, the activation function uses ReLU to remove the Softmax classification layer, and the loss function is cross entropy.

HNED structure: We regard vessel segmentation as an edge detection problem, and we use a network based on deep supervision to obtain four vessel probability maps of a shallow FCN network. That is, we add a softmax classifier after C2, C4, C6, and C8 respectively, and display the information of the hidden layer in the form of a retinal blood vessel probability map through a deep supervision network targeting the standard segmentation results, which are called side outputs respectively. 1. Side output 2, side output 3, and side output 4, to realize the learning of multi-scale feature maps.

Since the low-level feature resolution of the FCN network is relatively high, and the high-level information reflects stronger semantic information, it is very robust for the classification of blood vessels in some lesions and other areas of the retinal image, but finally the same size as the input sample is obtained. However, the output will lose a lot of smaller targets and local details. Therefore, the present invention learns rich multi-layer information expression under the condition of deep supervision by using the shallow retinal blood vessel information with the method of edge detection HNED, which solves the problem to a large extent. The problem of blurring the edge of the target is solved. On this basis, we fuse the four side outputs with the final output layer to form the network structure of FCN-HNED, as shown in Figure 4, if the input image size is 512×512, after C1 and C2 are both 64 A 3×3 filter obtains 64 feature maps. By padding the original image with zeros, the C1 and C2 feature maps maintain a size of 512×512. After downsampling, the feature maps are doubled and reach the bottom C9 and C10. When , the size of 1024 feature maps is 32×32, and the subsequent convolution implementation is similar to the previous one, and the implementation of upsampling is bilinear interpolation. The network structure complementarily fuses the four side output vessel probability maps of the shallow layer information with the output layer vessel probability map of the FCN network, and obtains a better feature map that is closer to the target sample through training, which plays a role in the refinement of the segmented blood vessels. to such an extent that no subsequent dedicated refinement step is required for retinal vessel refinement.

Fusion process: In order to directly use the side output probability map and the output probability map after FCN upsampling, we fuse them to get: Among them, σ( ) represents the sigmoid function, Represents the mth side output, h _m and h are the fusion weights of the four side outputs and the final output of the FCN, respectively, and the initial fusion weights are set to 1/5. The loss function for weighted fusion is:

Among them, Y represents the standard blood vessel segmentation map, which is the ground truth, and Dist(·,·) represents the distance between the fused probability map and the standard blood vessel segmentation map, that is, the degree of difference. Adjusting the weights by learning is gradually approaching convergence. We Minimize its loss function via SDG (gradient descent).

Training: After the FCN-HNED network is constructed, the network training can be carried out to carry out the automatic feature extraction and learning process of the image, which is divided into two steps: the first step is to manually select some 1280 pictures that are more intuitive, and first to construct the image in this paper. The model is trained, and 128 images are input in each generation. After the model converges, the model parameters are saved, because the content of these 1280 images is relatively intuitive and simple, the semantic information of blood vessels and non-vascular is relatively clear, and the convergence speed of the model is relatively fast; the second In the first step, the model is retrained on the full training set, but the initial value of the model parameters is the parameter obtained in the first step, which greatly reduces the training time of the model and speeds up the convergence of the overall model.

Training: After each image training data is calculated layer by layer through the convolutional neural network algorithm, a fused blood vessel probability map is output, and the error between the probability map and the corresponding category of each pixel in the standard map is calculated. According to the minimum error criterion, the parameters of each layer in the deep convolutional neural network constructed by error calculation are corrected layer by layer. When the error gradually decreases and becomes stable, it is considered that the network has converged, the training is over, and the required detection model is generated.

Test: Input the CLAHE map and Gaussian matching filter map of each retinal fundus image green channel map to the trained network for testing, and the fused retinal vessel segmentation map is called with right with Weighted averaging is performed to obtain more blood vessel information, and the final retinal blood vessel segmentation probability map is also obtained.

4 post-processing

Binarize the comprehensive retinal vessel probability map to obtain a segmentation map, which presents a binary map consistent with expert segmentation. By evaluating the parameters of the segmentation results, an accuracy rate of more than 96% is obtained, as shown in Figure 5.

Claims (1)

1. the Segmentation Method of Retinal Blood Vessels being combined with conventional method based on deep learning, it is characterised in that including following step Suddenly：

(1), pre-process

Tri- passage Green passages of RGB to colored retinal images are extracted, and are normalized, to normalization Retinal images afterwards carry out CLAHE treatment, and the retinal vessel after CLAHE is processed carries out Gauss matched filtering treatment, Retinal vessel figure and Gauss matched filtering figure after CLAHE is processed are all as the sample of training；

(2), data amplification and structure training sample

Data expand mode：

1) pretreated image is carried out into inferior translation on left and right and is respectively 20 pixels, realize the translation invariant of e-learning Property；

2) image after 1) processing carries out 45 ° respectively, and 90 °, 125 °, 180 ° of rotation intercepts maximum rectangle therein；

3) image set after 2) processing chooses wherein 25% and carries out 3 × 3 and 5 × 5 medium filtering fuzzy operation respectively；

4) image set after 3) processing takes 25% and enters row stochastic elastic deformation；

5) image after 4) processing carries out 50% and 75% scaling treatment, so that amplification data；

Carry out same treatment for the expert standard drawing ground truth that retinal vessel is split, so as to a pair of sample 1 Should；

(3), FCN-HNED network structions

The network of structure is：

Input layer, two convolutional layers (C1, C2), the first down-sampled layer (pool1), two convolutional layers (C3, C4), second is down-sampled Layer (pool2), two convolutional layers (C5, C6), the 3rd down-sampled layer (pool3), two convolutional layers (C7, C8), the 4th is down-sampled Layer (pool4), two convolutional layers (C9, C10), the first up-sampling layer (U1), two convolutional layers (C11, C12), the second up-sampling Layer (U2), two convolutional layers (C13, C14), the 3rd up-sampling layer (U3), two convolutional layers (C15, C16), the 4th up-sampling layer (U4), two convolutional layers (C17, C18), destination layer (output layer)；Form U-shaped depth network architecture symmetrical before and after；

By C2, a softmax grader is added after C4, C6, C8 layer respectively, so as to by the information of hidden layer by ground Truth be label in the case of study obtain retinal vessel probability graph, be referred to as side output 1, side output 2, side output 3, Side output 4；Four side outputs are merged with last output layer, so as to form the network structure of FCN-HNED；

Training：The training of network can be carried out after FCN-HNED network structions well to carry out to the Automatic Feature Extraction of image and Learning process, 128 images of per generation input, stops after network convergence；

Test：The CLAHE figures and Gauss matched filtering figure of every retinal images green channel figure are separately input to train Good network is tested, and the retinal vessel segmentation figure for respectively obtaining fusion is referred to asWithIt is rightWithCarry out Weighted average obtains last retinal vessel segmentation probability graph；

(4) post-process

Binaryzation is carried out to obtaining retinal vessel probability graph in test and obtains segmentation figure.