CN110751636B - Fundus image retinal arteriosclerosis detection method based on improved coding and decoding network - Google Patents

Abstract

The invention provides a fundus image retinal arteriosclerosis detection method based on an improved codec network, which comprises the following steps: 1) Collecting fundus images which are diagnosed as retinal arteriosclerosis by ophthalmologists, fitting lesion blood vessels in the fundus images, and calculating and counting reflex parameters to determine a retinal arteriosclerosis detection threshold; 2) The coding and decoding network is improved by utilizing a Inception ResnetV module and a residual attention mechanism module, and the method is applied to segmentation of arterial blood vessels and arterial reflective bands; 3) Screening an effective area, sampling the effective area, fitting by using four Gaussian models to obtain a vascular gray distribution curve, calculating a reflection parameter according to a fitting result, comparing the reflection parameter with a threshold value, and judging whether the patient suffers from retinal arteriosclerosis. The invention determines the quantitative detection threshold of the retinal arteriosclerosis, solves the problem that the traditional method can not accurately divide the artery and vein blood vessel of the fundus and the reflective belt of the artery by utilizing the deep learning technology, and completes the detection of the retinal arteriosclerosis.

Description

A detection method for retinal arteriosclerosis in fundus images based on improved codec network

technical field

The invention relates to a method for detecting retinal arteriosclerosis in fundus images based on an improved encoding and decoding network, which solves the problem that the traditional method cannot accurately segment fundus arteriovenous vessels and arterial reflective bands, and reduces the reflection of light and other fundus tissue characteristics on arteries The interference brought by band segmentation improves the transmission efficiency of feature and gradient information, realizes high-accuracy segmentation, determines the detection threshold of retinal arteriosclerosis, and realizes high-accuracy detection of retinal arteriosclerosis. It belongs to the fields of image processing, deep learning and medical imaging.

Background technique

Retinal arteriosclerosis is related to the degree of systemic arteriosclerosis. Since the fundus blood vessels are the only blood vessels that can be directly observed non-invasively, regular detection of retinal arteriosclerosis can help understand the degree of systemic arteriosclerosis, and assist doctors to intervene in advance for the continuous aggravation of systemic arteriosclerosis. Retinal arteriosclerosis will affect the diameter of the fundus arteries and the degree of reflection of the arterial reflective tape. At present, hospitals mainly use fundus cameras to observe the condition of human eyes. When arteriosclerosis occurs, the reflective band of the artery in the fundus image will widen, and the color of the blood column will turn into metallic bright copper; when the arteriosclerosis continues to aggravate, the blood vessels will appear white and silvery.

Clinically, the doctor observes the patient's fundus image based on experience, and judges whether the patient has arteriosclerosis according to whether the width of the arterial reflective band in the picture is increased compared with the width of the arterial vessel, and whether the degree of reflection of the arterial reflective band is enhanced. The reflective tape width ratio and gray scale ratio are used as the basis for arteriosclerosis detection. Since the detection of retinal arteriosclerosis requires gray-scale fitting of arterial vessels and arterial reflective bands, it is necessary to extract the arterial vessels and arterial reflective bands in the fundus image. Although the segmentation of blood vessels in fundus images has been widely concerned, there are few studies on the segmentation of arteries and arterial reflective bands. At the same time, because the shape and trend of veins are similar to arteries, when the quality of fundus images is not good, it may interfere with the segmentation of arteries, and the shape of arterial reflective tape is elongated, which is easily affected by surrounding light , so the segmentation of arterial vessels and arterial reflective bands is slightly more difficult.

In view of the above problems, the present invention proposes a method for detecting retinal arteriosclerosis in fundus images based on an improved codec network, which can be used for large-scale screening of retinal arteriosclerosis.

Contents of the invention

The invention proposes a method for detecting retinal arteriosclerosis in fundus images based on an improved codec network. First, collect the fundus images diagnosed as retinal arteriosclerosis by ophthalmologists, take a sampling line perpendicular to the arteries and arterial reflective bands in the fundus image, and use the four-segment Gaussian model to fit the pixels on the sampling line. Obtain the gray distribution curve of the blood vessel cross-section, and then calculate the reflection parameter-bandwidth ratio and gray ratio to determine the quantitative detection threshold of retinal arteriosclerosis; after that, the arteriovenous vessels and arterial reflective bands in the fundus image training set were respectively Annotation can reduce the interference caused by similar characteristics of arteries and veins; based on the improved codec network, extract multi-scale feature information with the help of 4 Inception Resnet V2 modules in the coding part, and add residual attention after each feature extraction module The force mechanism module is used to enhance target feature information and improve network performance. In the decoding part, four upsampling layers are used to generate sparse feature maps, and then four convolution layers are used to generate dense feature maps, and each pixel is classified through SoftMax; After that, the Canny edge detection operator is used to find all the connected domains of the reflective belt and screen out the largest connected domain, and three sampling straight lines perpendicular to the area are taken; for the three groups of sampling pixels, the four-segment Gaussian model is used to fit them respectively, and the obtained The gray distribution curves of the three cross-sections of blood vessels; finally, according to the fitting results, the reflective parameters—gray ratio and bandwidth ratio—of the three groups of arterial vessels and arterial reflective tapes were calculated. Comparing the three sets of reflective parameters with the detection threshold, there are three situations: (1) When the three sets of reflective parameters are all less than the threshold, it is judged that there is no retinal arteriosclerosis in the fundus; (2) When one of the three sets of reflective parameters has a If the parameter is greater than the threshold, it is judged that there is retinal arteriosclerosis in the fundus; (3) when other conditions occur, it is judged that the fundus is suspected of retinal arteriosclerosis.

Realize the technical scheme of the present invention, comprise the following steps:

Step 1: Collect fundus images that have been diagnosed as retinal arteriosclerosis by ophthalmologists, take a sampling line perpendicular to the arteries and arterial reflective bands in the fundus image, and use a four-segment Gaussian model to fit the pixels on the sampling line , to obtain the gray distribution curve of the blood vessel cross-section, and then calculate the reflection parameter-bandwidth ratio and gray ratio to determine the quantitative detection threshold of retinal arteriosclerosis;

Step 2: mark the arteriovenous vessels and arterial reflective tapes in the fundus image with different colors, and then divide all the fundus images into 224×224 image blocks according to the input requirements of the network, and use them as the training set of the improved codec network;

Step 3: Use the improved codec network to segment the arterial vessels and arterial reflective bands of the fundus image, and screen to obtain the effective area of the arterial reflective bands for subsequent detection;

Step 4: Take three sampling straight lines perpendicular to the arterial vessels and arterial reflective tape in the effective area, and use four-segment Gaussian model to fit the pixel points on the sampling straight lines to obtain gray distribution curves of three blood vessel cross-sections respectively , according to the fitting results, the boundary coordinates and average gray value of the three groups of arterial vessels and arterial reflective tapes were calculated respectively;

Step 5: Calculate the reflective parameters of arterial vessels and arterial reflective bands in the three groups of fundus images—bandwidth ratio and gray ratio, and compare with the quantitative detection threshold of retinal arteriosclerosis to determine whether the patient has retinal arteriosclerosis.

Compared with prior art, the beneficial effect of the present invention is:

By using deep learning technology, it solves the problem that traditional methods cannot accurately segment fundus arteriovenous vessels and arterial reflective bands, and reduces the interference caused by light and other fundus tissue characteristics on the segmentation of arterial reflective bands, and improves the accuracy of features and gradient information. The transmission efficiency realizes the accurate segmentation of arterial vessels and arterial reflective bands, determines the quantitative detection threshold of retinal arteriosclerosis, and realizes the detection of retinal arteriosclerosis with high accuracy.

Description of drawings

The overall framework schematic diagram of Fig. 1 the present invention;

The structure of Fig. 2 improved codec network;

Figure 3 The structure of the Inception Resnet V2 module;

Figure 4 The structure of the residual attention mechanism module;

Figure 5 Example of the screening process of the arterial reflective tape to segment the effective area;

Fig. 6 An example of the gray scale distribution of arterial blood vessels and arterial reflective tape pixels;

Figure 7 Gaussian fitting examples of arterial vessels and arterial reflective tape.

Detailed ways

The present invention will be further described in detail below in conjunction with specific embodiments.

The overall framework schematic diagram of the present invention is shown in Figure 1, at first, collect the fundus image that the ophthalmology expert has diagnosed as retinal arteriosclerosis, take a sampling straight line perpendicular to the arterial vessel and arterial reflective band in the fundus image, for the sampling straight line Pixel points are fitted with a four-segment Gaussian model to obtain the gray distribution curve of the blood vessel cross-section, and then the reflection parameter-bandwidth ratio and gray ratio are calculated to determine the quantitative detection threshold of retinal arteriosclerosis; The image database of arteriovenous vessels and arterial reflective tapes of fundus images, so the fundus images of the hospital are collected and the ophthalmologists use the labeling tool to manually mark the training samples; the improved codec network is extracted with the help of 4 Inception Resnet V2 modules in the coding part Multi-scale feature information, and a residual attention mechanism module is added after each feature extraction module to enhance target feature information and improve network performance. In the decoding part, four upsampling layers are used to generate sparse feature maps, and then through four volumes The dense feature map is generated by lamination, and the classification of each pixel is realized by SoftMax, and then applied to the segmentation of arterial vessels and arterial reflective bands in fundus images; the Canny operator is used to detect the edge of the arterial reflective band, and the maximum connected domain of the arterial reflective band As the detection effective area, three sampling straight lines perpendicular to the arterial vessels in the effective area were taken; for the sampling pixel points, four segments of Gaussian functions were used to fit them respectively, and three gray distribution curves of blood vessel cross-sections were obtained, according to the fitting results The reflective parameters of three groups of arterial vessels and arterial reflective tapes were calculated, and compared with the detection threshold, there were three situations: (1) When the three groups of reflective parameters were all less than the threshold, it was judged that there was no retinal arteriosclerosis in the fundus; (2) When one of the three sets of reflection parameters is greater than the threshold, it is judged that the fundus has retinal arteriosclerosis; (3) when other conditions occur, it is judged that the fundus is suspected of retinal arteriosclerosis.

The specific implementation process of the technical solution of the present invention will be described below in conjunction with the accompanying drawings.

1. Determination of Retinal Arteriosclerosis Detection Threshold

Collect 30 fundus images that have been diagnosed as retinal arteriosclerosis by ophthalmologists, and take a sampling line perpendicular to the arteries and arterial reflective bands in the fundus image, and use a four-segment Gaussian model to fit the pixels on the sampling line , to obtain the gray distribution curve of the blood vessel cross-section, and then calculate the reflection parameter-bandwidth ratio and gray ratio to determine the quantitative detection threshold of retinal arteriosclerosis: bandwidth ratio and the gray ratio

Compared with the quantitative detection threshold, there are three situations: (1) when the three groups of reflective parameters are all less than the threshold, it is judged that there is no retinal arteriosclerosis in the fundus; (2) when one of the three groups of reflective parameters is greater than the threshold, then It is judged that there is retinal arteriosclerosis in the fundus; (3) when other conditions occur, it is judged that the fundus is suspected of retinal arteriosclerosis.

2. Subjects

The training set used in the present invention includes 918 training samples of 224×244 and 168 testing samples. Since the characteristics of arteries and veins are similar, and the shape of the arterial reflective tape is slender, it is easy to be misdetected or missed. However, because there is no public database of fundus arteriovenous blood vessels and arterial reflective tapes, it is necessary to use a graphical annotation tool Manually mark arteriovenous vessels and arterial reflective tapes in the training samples.

3. Improved codec network

The main body of the improved codec network used in the present invention is an encoding-decoding network. The feature extraction structure of the encoding part is similar to the VGG16 network, mainly composed of 4 convolutional blocks, which mainly include convolutional layers, batch normalization, activation functions, and pooling layers. At the same time, with the help of the Inception Resnet V2 module, the encoding part has a multi-scale input structure. The convolutional layer in the Inception Resnet V2 module uses a size of 1×1, 3×3, 1×3, 3×1, 1×7 and The purpose of the 7×1 convolution kernel is to expand the width of the network and extract feature information of different scales of the image. The encoding part uses maximum pooling for downsampling to retain salient features, reduce feature dimensions, and add padding to preserve boundary information. In addition, a residual attention mechanism module is added after the maximum pooling layer, so that the network can select the focused position and enhance the feature representation at this position. The main structure of the decoding part is similar to the decoding structure of the SegNet network. It uses 4 upsampling blocks, including convolution layer, activation function layer, batch normalization layer and upsampling layer, and adds a SoftMax classifier at the end of the decoding part to realize The segmentation of the overall network and the structure of the improved codec network are shown in Figure 2.

(1) Coding part

From the image input layer to the last pooling layer belongs to the encoding part. In this stage, the InceptionResnet V2 module is mainly used to expand the network width to extract multi-scale feature information. The residual attention mechanism module focuses the network on the target area and enhances network performance. First, five Inception Resnet V2-A modules are used to extract features from 224×224 images. The convolution kernels of this module include 1×1 and 3×3. The 1×1 convolution operation is to limit the number of input channels. , and two consecutive 3×3 convolution operations can be regarded as feature extraction using a 5×5 convolution kernel, and then use a 1×1 convolution operation to make the input and output have the same dimension, and use a residual The feature is fused in the way of poor connection; the size of the feature map is reduced to 112×112 after the maximum pooling layer, and it is used as the input of the Stage1 module of the residual attention mechanism, which is a spatial attention module (Spatial Attention ), using L2 regularization to constrain all channels at each position, and finally output an attention map with consistent spatial dimensions; then, connect 5 Inception Resnet V2-B modules, the convolution kernels of which are 1×1 , 1×7 and 7×1, the last two convolution operations are equivalent to using a 7×7 convolution kernel, and the extracted features are fused and sent to the maximum pooling layer to obtain a 56×56 feature map; The 56×56 feature map is sent to the residual attention mechanism module Stage2. The Stage2 module belongs to the channel attention module, similar to the SENet network that constrains all feature values on each channel, so that the output length is the same as the number of channels. One-dimensional vector As privilege weighting; after that, connect 5 Inception Resnet V2-C modules, the convolution kernels used by this module are 1×1, 1×3 and 3×1 respectively, and the last two convolution operations are equivalent to using 3×3 The convolution kernel, the feature map is pooled to the maximum, and the feature map of 28×28 is obtained; the residual attention mechanism Stage3 module is connected, which belongs to the attention concentration module, and is introduced by using Sigmoid for each channel and each space Non-linear factors; finally, connect 5 InceptionResnet V2-C modules again to extract high-level features, and connect the maximum pooling layer to reduce the dimension of the feature map to obtain a 14×14 feature map. Since the residual attention mechanism is to focus on the target in a larger feature map, when the size of the feature map is small enough, there is no need to add a residual attention mechanism.

(2) Decoding part

The decoding part starts from the upsampling layer after the maximum pooling layer to the end of the SoftMax layer. The decoding part mainly decodes and transmits the information of the coding part layer by layer through the upsampling layer and the convolutional layer. The 14×14 feature map passes through the upsampling layer (Upsample) to generate a sparse feature map, and then undergoes a 3×3 convolution operation to generate a 28×28 dense feature map, and performs four upsampling and convolution in the same way operation, and finally send the high-dimensional features of the output to SoftMax for the classification of each element

3.1 Inception Resnet V2 module

In semantic segmentation, since objects occupy different sizes in different images, there is a huge difference in location information. When the information distribution is balanced, the image is suitable for a larger convolution kernel; when the information distribution is concentrated, the image is suitable for a smaller convolution kernel, so it is difficult to choose an appropriate convolution kernel size for the convolution operation. If the convolutional neural network simply stacks larger convolutional layers to extract information, then the network is extremely prone to overfitting, and the gradient information is difficult to transmit to the entire network, and the cost of computing resources will also increase. The Inception Resnet V2 module uses convolution kernels of multiple sizes at the same level to expand the network width and extract feature information of multiple scales. At the same time, in order to reduce information loss, three different types of Inception Resnet V2 modules were constructed, namely Inception Resnet V2-A, Inception Resnet V2-B and Inception Resnet V2-C, and the connection mode of the residual network was introduced to realize A more efficient feature extraction method is proposed. The structure of the Inception Resnet V2 module is shown in Figure 3.

3.2 Residual Attention Mechanism Module

Residual attention network is a convolutional neural network with stacked attention mechanism modules, which combines end-to-end training with the latest feed-forward network structure. As the number of network layers increases, the attention features of different attention mechanism modules will change adaptively, not only can select the focus position, capture different types of attention in the image, but also enhance the feature representation of the object at this position, thus Improve network performance. At the same time, the network uses a residual attention learning method, which can be used to train a deep convolutional neural network. Each residual attention module can be divided into two branches: the mask branch and the backbone branch, where the backbone branch uses the pre-activation residual module and the Inception module as the basic unit of the network; in the mask branch, the feature map Processing mainly includes forward downsampling and upsampling. Downsampling can quickly encode and obtain the global features of the feature map. The upsampling function is to combine the extracted global high-dimensional features with non-downsampled features, so that high- and low-dimensional features can be better combined, and the output dimension is consistent. Attention to the feature map, and then use the dot multiplication operation to combine the feature maps of the two branches to obtain the final output feature map. The output result is shown in the following formula:

H _i,c (x)=(1+M _i,c (x))×F _i,c (x)

Among them, M _{i, c} (x) represents the feature output map of the mask branch, F _{i, c} (x) is the residual result between the output and the input, which is learned and fitted by a deep convolutional neural network, i Indicates the range of spatial locations, and c indicates the channel index.

The residual attention mechanism module is divided into three types, including spatial attention, channel attention, and attention concentration. The structure of the residual attention mechanism module is shown in Figure 4.

4. Effective area screening of arterial reflective tape

In order to reduce the workload of subsequent arteriosclerosis detection, it is necessary to screen the effective area of arterial reflective tape segmentation for the segmentation results. An example of the screening process of the effective region of arterial reflective tape segmentation is shown in Figure 5, where Figure 5(a) is the arterial reflective tape extracted from the segmentation result; Figure 5(b) is the result of binarization of the arterial reflective tape; 5(c) is the result of using the Canny operator to find the connected domain of the arterial reflective band; finally, the maximum connected domain of the arterial reflective band is obtained by screening, and the subsequent arterial vessels and arterial reflective bands to be detected are obtained, as shown in Figure 5(d).

5. Gaussian fitting of arterial vessels and arterial reflective tape

An example of the gray distribution of arterial vessels and arterial reflective tape pixels is shown in Figure 6. The upper curve in the figure is the radial gray distribution curve of arterial vessels, and the lower curve is the first-order derivative of the gray distribution curve. The gray distribution curve is extremely Small value points C and D are the boundaries of the arterial reflective belt, and M is the highest point of the gray distribution curve. Since the arterial reflective tape is located in the center of the arterial vessel, and its brightness is higher than that of the blood vessel, the radial gray distribution of the arterial vessel conforms to the distribution law of the Gaussian function, and the coordinates of the minimum values C and D can be determined by calculating the first derivative of the gray distribution curve . Therefore, the present invention proposes to use a four-segment Gaussian model to describe the gray distribution of arterial blood vessels and arterial reflective tape cross-sections. Arterial reflective tape border.

Four sections of Gaussian function expressions that the present invention adopts are:

Among them, a ₁ , a ₂ , a ₃ , and a ₄ are the peak values of the Gaussian curve, b ₁ , b ₂ , b ₃ , and b ₄ are the peak coordinates, and c ₁ , c ₂ , c ₃ , and c ₄ are the standard deviations.

In order to accurately describe the radial gray distribution of arteries, the present invention performs three samplings on the arteries to be tested, and performs Gaussian fitting respectively. An example of Gaussian fitting of arteries and arterial reflective tapes is shown in FIG. 7 . Fig. 7 (a) is the sampling of arterial blood vessel, and the present invention carries out four sections of Gaussian fittings to sampling pixel points by determining a sampling straight line perpendicular to blood vessels; Fig. 7 (b) uses four sections of Gaussian functions to sample pixel points The result of gray value fitting; Figure 7(c) is the first derivative of the gray value Gaussian fitting result, and the coordinates of the minimum point C and D can be calculated from the figure. The definition of each region in the fitting result of the gray distribution in Figure 7(b) is shown in Table 1: points A and B are arterial borders, A ^- and B ⁺ are retinal backgrounds, C and D are gray distribution poles The small point is the border of the arterial reflective band, CM and MD are the area of the arterial reflective band, and point M is the peak of the CD segment.

Definition of arterial vessels and arterial reflective tape areas in the example of gray scale distribution curve in Table 1

area area definition A≤x||x≥B <![CDATA[Background (A^-, B⁺)]]> A<x<C||D<x<B Arterial vessels (AC, DB) C<x≤M||M<x<D Arterial reflective tape (CM, MD) 

6. Detection of retinal arteriosclerosis

The reflective parameters of arteries and arterial reflective tapes include bandwidth ratio BR and gray scale ratio GR. According to the fitting results of arterial vessels and arterial reflective tapes, the coordinates of arterial vessels and arterial reflective tape boundaries are obtained, and then the width and average gray level of arterial blood vessels and arterial reflective tapes on the three straight line sampling points are respectively calculated, and finally the gray ratio and bandwidth are calculated. Compare. When retinal arteriosclerosis occurs, the arterial reflective band becomes wider and produces hyperreflection, resulting in an increase in the reflective parameters of the arterial vessel and arterial reflective band. Therefore, the set of values with the largest reflective parameter among the three sets of data is selected as the patient's arteriosclerosis detection. reflective parameters. The calculation formula of reflective parameters is as follows:

Bandwidth ratio:

Gray ratio:

Among them, X _A , X _B , X _C and X _D are the boundary coordinates of blood vessels and arterial reflective tapes, Y _i is the gray value corresponding to each point, n ₁ and n ₂ are the number of pixels of arterial reflective tapes and blood vessels respectively, X _D -X _C is the width of arterial reflective tape, X _B -X _A is the width of arteries, is the average gray value of the arterial reflective tape, is the average gray value of arterial vessels.

The present invention uses 53 fundus images provided by the hospital as test images, including 16 fundus images of normal fundus, 14 fundus images of suspected retinal arteriosclerosis, and 23 fundus images of arteriosclerosis. The present invention is used to detect retinal arteriosclerosis on the above-mentioned fundus image. By comparing the three sets of reflective parameters of the fundus image to be tested with the detection threshold, there are three situations: (1) when the three sets of reflective parameters are all less than the threshold, it is judged that there is no retinal arteriosclerosis in the fundus; (2) when the three sets of reflective parameters If one group of parameters among the parameters is greater than the threshold, it is judged that there is retinal arteriosclerosis in the fundus; (3) when other conditions occur, it is judged that the fundus is suspected of retinal arteriosclerosis. Among them, the detection accuracy rate of normal fundus images was 93.7%, the detection accuracy rate of fundus images suspected of having retinal arteriosclerosis was 92.8%, and the detection accuracy rate of fundus images suffering from retinal arteriosclerosis was 91.3%. was 92.6%. In summary, the present invention determines the detection threshold of retinal arteriosclerosis, and proposes a method for detecting retinal arteriosclerosis in fundus images based on an improved codec network, which can accurately segment retinal arteriovenous vessels and arterial reflective bands, and solves the problem of traditional methods. The problem that fundus arteriovenous vessels and arterial reflective bands cannot be accurately segmented, and the interference caused by light and other fundus tissues on the arterial reflective band segmentation is reduced, the transmission efficiency of feature and gradient information is improved, and high-accuracy segmentation is achieved. The detection of retinal arteriosclerosis was completed.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. It should be understood that the present invention is not limited to the implementation solutions described here. The purpose of these implementation solutions descriptions is to help those skilled in the art Those skilled in the art practice the present invention. Any person skilled in the art can easily carry out further improvement and perfection without departing from the spirit and scope of the present invention, so the present invention is only limited by the content and scope of the claims of the present invention, and it is intended to cover all Alternatives and equivalents within the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A fundus image retinal arteriosclerosis detection method based on an improved Inception Resnet V attention codec network comprises the following steps:

step 1: collecting fundus images which are diagnosed as retinal arteriosclerosis by ophthalmologists, taking a sampling straight line perpendicular to arterial blood vessels and arterial reflecting bands in the fundus images, fitting pixel points on the sampling straight line by using four Gaussian models to obtain gray level distribution curves of blood vessel cross sections, then calculating reflecting parameters-bandwidth ratio and gray level ratio, and determining a retinal arteriosclerosis quantitative detection threshold;

step 2: marking arteriovenous blood vessels and arterial reflection bands in fundus images respectively by using different colors, dividing all fundus images into 224×224 image blocks according to input requirements of a network, and taking the image blocks as a training set of an improved Inception Resnet V attention coding and decoding network;

step 3: extracting multi-scale characteristic information by means of 4 Inception Resnet V modules in a convolution block of an improved Inception Resnet V attention codec network coding part, and adding a residual attention mechanism module after each characteristic extraction module to strengthen target characteristics and improve network performance; generating a sparse feature map by utilizing four upsampling layers in a decoding part of the improved Inception Resnet V attention codec network, generating a dense feature map by using four convolution layers, realizing classification of each pixel by softMax, then applying the classification to segmentation of fundus image artery and vein vessels and artery reflection bands, and screening to obtain an artery reflection band effective region for subsequent detection;

step 4; three sampling straight lines perpendicular to the arterial blood vessel and the arterial reflective band are taken in the effective area, fitting is carried out on pixel points on the sampling straight lines by using four Gaussian models to respectively obtain gray level distribution curves of the cross sections of the 3 blood vessels, and boundary coordinates and average gray level values of the arterial blood vessel and the arterial reflective band of the 3 groups are respectively calculated according to fitting results;

step 5; and respectively calculating reflection parameters-bandwidth ratio and gray scale ratio of artery blood vessels and artery reflection bands of the 3 groups of fundus images, comparing with a quantitative detection threshold value of retinal arteriosclerosis, and judging whether the patient suffers from retinal arteriosclerosis.

2. The method for detecting retinal arteriosclerosis in fundus images based on the improved Inception Resnet V attention codec network according to claim 1, wherein in step 1, fundus images which are diagnosed as retinal arteriosclerosis by ophthalmologists are collected, a sampling straight line perpendicular to arterial blood vessels and arterial reflective bands in the fundus images is taken, pixel points on the sampling straight line are fitted by using a four-segment Gaussian model, a gray level distribution curve of a blood vessel cross section is obtained, and then calculation of reflective parameter-bandwidth ratio and gray level ratio is carried out to determine a retinal arteriosclerosis quantitative detection threshold.

3. The method for detecting retinal arteriosclerosis in fundus images based on the improved Inception Resnet V attention codec network according to claim 1, wherein in the step 2, the arteriovenous vessels and the arterial reflection bands in the fundus images are marked by different colors respectively, so that the interference of venous vessels on segmentation is reduced.

4. The method for detecting retinal arteriosclerosis by fundus images based on the improved Inception Resnet V attention codec network according to claim 1, wherein in the step 4, three sampling straight lines perpendicular to arterial blood vessels and arterial reflective bands are taken in an effective area, pixel points on the sampling straight lines are fitted by using four Gaussian models, gray level distribution curves of cross sections of the 3 blood vessels are obtained respectively, and boundary coordinates and average gray level values of the arterial blood vessels and arterial reflective bands of the 3 groups are calculated respectively according to fitting results.

5. The method for detecting retinal arteriosclerosis in fundus images based on the improved Inception Resnet V attention codec network according to claim 1, wherein in step 5, reflection parameters of 3 groups of arterial blood vessels and arterial reflection bands are calculated respectively, and three conditions exist for comparing the reflection parameters with detection thresholds: (1) When the three groups of reflection parameters are smaller than the threshold value, judging that retinal arteriosclerosis does not exist in the fundus; (2) When one of the three groups of reflection parameters is larger than a threshold value, judging that retinal arteriosclerosis exists in the fundus; (3) When other conditions occur, the fundus is judged to be suspected retinal arteriosclerosis.

Images