A method, system and device for calculating retinal blood vessel curvature

Abstract

The present application relates to the field of intelligent medical treatment, and specifically to a method, system and device for calculating the curvature of retinal blood vessels. The method includes acquiring a fundus image; performing blood vessel segmentation on the fundus image to obtain arterial blood vessels and venous blood vessels; performing skeleton extraction on the arterial blood vessels and venous blood vessels to obtain a skeleton image; screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image; and calculating the curvature of the blood vessels based on the screened image. The curvature of the retinal arterial blood vessels and venous blood vessels calculated by the present application can effectively predict the risk of cardiovascular disease and has good clinical value.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical treatment, and specifically to a method, system, device and computer-readable storage medium for calculating retinal vascular curvature. Background Art

[0002] The retina is currently the only part of the body where microvessels can be directly observed non-invasively. Retinal microvascular abnormalities are associated with hypertension, atherosclerosis, and cardiovascular disease. In addition, retinal artery stenosis may reflect the severity of cardiovascular disease complications and serve as an independent risk factor for cardiovascular disease. Therefore, identifying retinal vascular abnormalities and quantifying vascular curvature may play an important role in the long-term monitoring of cardiovascular disease.

[0003] Over the past few decades, researchers have developed various methods to classify retinal vascular abnormalities caused by hypertension. These classification systems are mainly based on global arterial constriction, arteriovenous compression, arterial lining, changes in arterial light reflection, and hypertensive retinopathy (including microaneurysms, intraretinal hemorrhages, cotton-wool spots, hard exudates, and optic disc edema). Although these classifications are easy to perform during fundus examination, they are relatively subjective and their reproducibility is questionable. In addition, changes in the curvature of retinal vessels have been associated with hypertension, coronary artery disease, and stroke. Therefore, a program that can automatically measure the curvature of arteries and venous vessels is needed. Summary of the invention

[0004] Artificial intelligence, especially deep learning, has achieved great success in natural image classification tasks, providing a variety of texture and semantic information that is different from manually designed image features. Recent studies have shown that deep learning can improve the traditional analysis of vascular structure in fundus images. Therefore, in order to solve the above problems, the present invention uses deep learning to propose a method for calculating retinal vascular curvature, which specifically includes:

[0005] Acquire fundus images;

[0006] Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels;

[0007] Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0008] Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0009] The curvature of the blood vessel is calculated based on the screened image.

[0010] Furthermore, the screening is to identify the blood vessel bifurcation point through the graph data structure, and then screen out the blood vessels containing the blood vessel bifurcation point to obtain the screened image;

[0011] Preferably, the graph data structure adopts an undirected graph for recognition, and the recognition constructs a graph node for each pixel point of the skeleton image, and then constructs an edge for adjacent pixel points to form an undirected graph; the degree of the node of the undirected graph is judged, and when the degree of the node of the undirected graph is greater than a preset threshold, it is determined to be a bifurcation point.

[0012] The screening also includes screening out too short or too thin blood vessels from the skeleton image, obtaining primary screening blood vessels after screening out too short or too thin blood vessels from the skeleton image, identifying blood vessel bifurcation points from the primary screening blood vessels, and screening out blood vessels containing blood vessel bifurcation points to obtain a screened image;

[0013] Preferably, the specific steps of screening out the too short or too thin blood vessels are:

[0014] The first step is to binarize the skeleton image to obtain a binary image;

[0015] The second step is to extract the centerline of the binary image to obtain the centerline, and obtain the distance from each pixel on the blood vessel to the centerline based on the centerline;

[0016] The third step is to calculate the average value of the distances to obtain an average distance and use the average distance as the blood vessel width;

[0017] Step 4: Delete the blood vessels whose width is less than a preset threshold, and delete the blood vessels whose length is less than m pixels to obtain the initial screening blood vessels, where m is a positive number less than or equal to 20;

[0018] Optionally, the preset threshold is to sort all the calculated blood vessel widths, select the first n blood vessel widths and calculate the average to obtain a width mean, where n is a natural number greater than 2; and use the value obtained by subtracting k times the standard deviation from the width mean as the threshold, where k is a natural number greater than 1;

[0019] Preferably, the step of screening out blood vessels that are too short or too thin also includes selecting a target area, after obtaining the distance from each pixel on the blood vessel to the center line, taking the center of the optic disc as the center and an area formed by c times the diameter of the optic disc as the target area, and calculating the blood vessels in the target area to obtain the blood vessel width, where c is a natural number greater than or equal to 1.

[0020] The segmentation is performed by one or more of the following segmentation models: U-net, FCN, SegNet, DeepLab v3, Faster R-CNN, PSPnet, ESPNet, Mask R-CNN;

[0021] Preferably, the segmentation is performed by a U-net network framework and a boundary enhancement module, edge attention mapping features are generated by edge detection of the fundus image, and the edge attention mapping features are input to a convolutional layer of an encoding module of the U-net network framework to enhance feature representation;

[0022] Preferably, the segmentation further includes optic disc segmentation, performing optic disc segmentation on the fundus image to obtain an optic disc; and selecting a target area based on the optic disc;

[0023] Preferably, the method further includes data preprocessing, which includes one or more of the following: dark channel prior assumption, image cropping and filling, grayscale, geometric transformation, binarization, and histogram equalization.

[0024] The calculation is performed by fitting the coordinates of the pixel points in the screened image into a curve to obtain an approximate curve of the blood vessel, and then performing curvature calculation based on the curve to obtain the curvature of the blood vessel;

[0025] Preferably, the pixel point coordinates are fitted as a curve by fitting a cubic polynomial;

[0026] Preferably, the pixel coordinates in the screened image are fitted into a curve and replaced by: connecting two ends of the blood vessel in the screened image to obtain a connecting line midpoint and using it as a rotation center, rotating the blood vessel based on the rotation center so that the connecting line midpoint is parallel to the horizontal axis in the horizontal axis coordinate to obtain the rotated blood vessel, mapping the rotated blood vessel to the horizontal and vertical coordinates to obtain updated pixel coordinates, and fitting the updated pixel coordinates into a curve to obtain a blood vessel approximate curve;

[0027] Optionally, the curve further includes one or more of the following calculations: arc length calculation of the curve, integral calculation of the square of the curvature of the curve, cTORT calculation;

[0028] Optionally, the calculation formula of cTORT is:

[0029]

[0030] Where tsc(C) represents the integral of the square of the curvature, s(C) represents the arc length of the curve, and C represents the curve.

[0031] The object of the present invention is to provide a method for calculating the diameter of retinal blood vessels, comprising:

[0032] The method implements any one of the methods for calculating the curvature of retinal blood vessels, and replaces the step of calculating the curvature of the blood vessel based on the screened image with: calculating the diameter of the blood vessel based on the screened image;

[0033] Optionally, the calculation of the diameter includes one or more of the following parameters: vessel diameter, average vessel thickness, CRAE, CRVE, AVR;

[0034] Preferably, the step of identifying the blood vessel bifurcation point is replaced by: first extracting the center line of the skeleton image and obtaining the topological structure of blood vessel distribution based on the center line, and then identifying the bifurcation point in the topological structure by a sliding window method;

[0035] Preferably, the step of removing the blood vessels containing the blood vessel bifurcation point is replaced by: taking the blood vessels near the center of the optic disc in the target area as the main blood vessels, wherein the blood vessels containing the blood vessel bifurcation point in the main blood vessels are removed, and the bifurcated blood vessels are removed and the main trunk is retained; the blood vessels whose bifurcation points are less than d pixels away from the boundary of the target area are removed from the main trunk and the bifurcated blood vessels are retained, where d is a natural number greater than 1; and finally, the image after the removal is obtained;

[0036] Preferably, the average blood vessel thickness is calculated by first calculating the blood vessel thickness and then averaging the blood vessel thickness to obtain the average blood vessel thickness;

[0037] Preferably, the blood vessel thickness is the distance between the center line and the blood vessel boundary;

[0038] Preferably, the blood vessel diameters in the screened images are sorted, and the first h blood vessel diameters are selected and averaged to obtain the average blood vessel diameter, where h is a natural number greater than 1.

[0039] The object of the present invention is to provide a method for predicting disease risk based on retinal blood vessels, comprising:

[0040] The method implements any one of the methods for calculating the curvature of retinal blood vessels to obtain the curvature of retinal blood vessels;

[0041] The method implements any one of the methods for calculating the diameter of retinal blood vessels to obtain the diameter of retinal blood vessels;

[0042] Preferably, the method implements any one of the methods for calculating the curvature of the retinal blood vessels, and simultaneously calculates the diameter of the blood vessels based on the screened images to obtain the curvature of the retinal blood vessels and the diameter of the retinal blood vessels;

[0043] The curvature of the retinal blood vessels and the diameter of the retinal blood vessels are input into a prediction model to obtain a disease prediction result.

[0044] The object of the present invention is to provide a system for calculating the curvature of retinal blood vessels, comprising:

[0045] Acquisition module: acquire fundus images;

[0046] Segmentation module: segment the fundus image to obtain arterial and venous vessels;

[0047] Extraction module: extracting skeletons of the arterial blood vessels and venous blood vessels to obtain skeleton images;

[0048] Screening module: Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0049] Calculation module: Calculate the curvature of the blood vessel based on the screened image.

[0050] The object of the present invention is to provide a system for calculating the curvature of retinal blood vessels, comprising:

[0051] The system executes the calculation method for the retinal vessel curvature.

[0052] The object of the present invention is to provide a system for calculating the diameter of retinal blood vessels, comprising:

[0053] Acquisition module: acquire fundus images;

[0054] Segmentation module: segment the fundus image to obtain arterial and venous vessels;

[0055] Extraction module: extracting skeletons of the arterial blood vessels and venous blood vessels to obtain skeleton images;

[0056] Screening module: Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0057] Calculation module: The diameter of the blood vessel is obtained by performing diameter calculation based on the screened image.

[0058] The object of the present invention is to provide a system for calculating the diameter of retinal blood vessels, comprising:

[0059] The system executes the method for calculating the retinal blood vessel diameter.

[0060] The object of the present invention is to provide a system for predicting disease risk based on retinal blood vessels, comprising:

[0061] The system executes and implements the method for predicting disease risk based on retinal blood vessels.

[0062] The object of the present invention is to provide a device for calculating the curvature of retinal blood vessels, comprising:

[0063] The device executes the method for calculating the curvature of the retinal blood vessels.

[0064] The object of the present invention is to provide a device for calculating the diameter of retinal blood vessels, comprising:

[0065] The device executes the method for calculating the retinal blood vessel diameter.

[0066] The object of the present invention is to provide a device for predicting disease risk based on retinal blood vessels, comprising:

[0067] The device executes the method for predicting disease risk based on retinal blood vessels.

[0068] The object of the present invention is to provide a computer-readable storage medium device, comprising:

[0069] A memory and a processor, wherein the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, the method for calculating the retinal vascular curvature is implemented.

[0070] The object of the present invention is to provide a computer-readable storage medium device, comprising:

[0071] A memory and a processor, wherein the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, the method for calculating the retinal blood vessel diameter is implemented.

[0072] The object of the present invention is to provide a computer-readable storage medium device, comprising:

[0073] A memory and a processor, wherein the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, the method for predicting disease risks based on retinal blood vessels is implemented.

[0074] Advantages of the present invention:

[0075] 1. An undirected graph is used to identify the bifurcation points of retinal blood vessels. The degree of the undirected graph can well describe the extensibility of the bifurcation point. By using the degree of the undirected graph to represent the direction of blood vessel extension, it is easier to identify the bifurcation points in the blood vessels and improve the efficiency of detecting bifurcation points.

[0076] 2. Use pixel rotation and remapping to obtain new pixel coordinates to ensure that the blood vessel pixels and the horizontal and vertical coordinates form a one-to-one mapping function relationship, thereby obtaining a more accurate curve when fitting the blood vessel curve and improving the accuracy of blood vessel curvature calculation.

[0077] 3. The calculation of the curvature and diameter of retinal blood vessels is proposed, and disease prediction is performed based on the curvature and diameter of the blood vessels, which increases the accuracy of disease prediction.

[0078] 4. The quality and consistency of data obtained from automated analysis of fundus images have been significantly improved, especially when processing fundus images of different sources and qualities, providing an effective tool for automated diagnosis of fundus lesions and MACE risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0080] Figure 1 A schematic diagram of a flow chart of a method for calculating retinal vessel curvature provided in an embodiment of the present invention;

[0081] Figure 2 A schematic diagram of a device provided by an embodiment of the present invention;

[0082] Figure 3 A schematic diagram of a segmentation model structure provided by an embodiment of the present invention;

[0083] Figure 4 The Kaplan-Meier curve of the cTORTa two-grouping and the risk of MACE provided in the embodiment of the present invention, B1: 0.06E-4 ~ 1E-4, B2: 1E-4 ~ 21.84E-4; cTORTa, curvature tortuosity arteriole, retinal artery curvature curvature;

[0084] Figure 5 The Kaplan-Meier curve of the two-group cTORTv and the risk of new MACE provided in the embodiment of the present invention, B1: 0 ~ 1.85E-4, B2: 1.85E-4 ~ 36.08E-4; cTORTv, curvature tortuosity venule, retinal venous curvature tortuosity. DETAILED DESCRIPTION

[0085] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0086] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0087] Figure 1 The schematic diagram for calculating the curvature of retinal blood vessels provided in the embodiment of the present invention specifically includes:

[0088] S101: Acquire fundus images;

[0089] In one embodiment, the fundus generally refers to the inside of the back of the eyeball, which is composed of the optic disc, macula, retina, and some blood vessels. It appears orange-red, and the fundus condition can usually be explored through imaging technology.

[0090] In one embodiment, data preprocessing is performed on the fundus image, and the preprocessing includes one or more of the following: dark channel prior assumption, image cropping and filling, grayscale, geometric transformation, binarization, and histogram equalization.

[0091] In a specific embodiment, the present invention is a prospective cohort study based on the Beijing community population. The research subjects are from the atherosclerosis cohort established by the Department of Cardiology of Peking University First Hospital in Gucheng Community and Pingguoyuan Community in Shijingshan District, Beijing. The age of the research subjects is ≥40 years old. The inclusion criteria of the research subjects in the present invention are: (1) participation in the 2014 follow-up; (2) complete retinal fundus image information in 2014, good image quality, and no missing retinal vascular tortuosity measurement values; (3) no history of CVD (myocardial infarction and / or stroke) at baseline in 2014. Finally, a total of 3585 research subjects were included.

[0092] In 2014, retinal photographs of both eyes were taken using a color digital retinal camera (Canon, Japan), with one eye centered on the optic disc (Zone 1 of the Diabetic Retinopathy Research Standard) and the other eye centered on the macula (Zone 2 of the Diabetic Retinopathy Research Standard). The fundus images of the right eye were selected to establish the deep learning model. After removing unclear fundus images, 3791 single-eye fundus images were finally used to establish the deep learning model.

[0093] In a specific embodiment, the data set contains fundus images from different devices and shooting conditions, providing a diverse analysis basis for this study. However, different device parameters, lighting conditions, and photographers' operating habits may significantly affect the appearance of fundus images, making automated analysis difficult, especially in multicenter studies. To enhance the consistency of images, a series of preprocessing steps are used for these image data from different sources, including using the "dark channel prior" assumption to optimize image quality under lighting conditions, unifying the appearance of fundus images from different data sources, and converting all image data into squares through image cropping and padding. The preprocessing steps unify the appearance of different types of fundus images, and the purpose of this is to ensure image consistency during subsequent analysis.

[0094] S102: performing blood vessel segmentation on the fundus image to obtain arterial vessels and venous vessels;

[0095] In one embodiment, the segmentation is performed by one or more of the following segmentation models: U-net, FCN, SegNet, DeepLab v3, Faster R-CNN, PSPnet, ESPNet, Mask R-CNN;

[0096] In one embodiment, the segmentation is performed by a U-net network framework and a boundary enhancement module, edge attention mapping features are generated by edge detection of the fundus image, and the edge attention mapping features are input to a convolutional layer of an encoding module of the U-net network framework to enhance feature representation;

[0097] In one embodiment, the U-shaped network framework includes an encoding module and a decoding module, each of which includes M units, where M is a natural number greater than 1, and M-1 units in the encoding module and M-1 units in the decoding module are connected to form a U shape by skip connections to form units with the same feature dimensions; the fundus image is input to the segmentation model and sequentially encoded by the units in the encoding module to obtain a feature map, and the feature map is sequentially decoded by the units in the decoding module to obtain segmented arterial vessels and venous vessels;

[0098] Optionally, each unit in the encoding module includes a convolution layer and a pooling layer, the fundus image is sequentially input to the first unit of the encoding module, and sequentially passes through the convolution layer and the pooling layer to obtain a feature map, the feature map is input to the next unit after downsampling, and this step is repeated until the Mth unit;

[0099] Optionally, each unit in the decoding module includes a convolution layer and a normalization layer, the feature map is sequentially input to the first unit in the decoding module, and sequentially passes through the convolution layer and the normalization layer to obtain a decoding feature, the decoding feature is input to the next unit after upsampling, and this step is repeated until the Mth unit;

[0100] Generate edge attention mapping features by performing edge detection on the fundus image, and input the edge attention mapping features into the convolution layer of the M-1 units of the encoding module to enhance feature representation;

[0101] In one embodiment, the segmentation further includes optic disc segmentation, performing optic disc segmentation on the fundus image to obtain an optic disc; and selecting a target area based on the optic disc;

[0102] In a specific embodiment, the present invention develops a deep learning model based on a U-shaped network structure, such as Figure 3 As shown, it is used for accurate segmentation of retinal blood vessels and classification of arteries and veins. The model extracts and reconstructs image features through encoding and decoding paths, and introduces a boundary enhancement module to further optimize the smoothness and accuracy of the segmentation results. On this basis, the present invention also implements automatic segmentation of the optic disc, which is accomplished through unlabeled learning methods and data set fusion technology, a step that is crucial to ensure the accuracy of the analysis.

[0103] In a specific embodiment, retinal blood vessel segmentation and arteriovenous classification: using custom software designed specifically for the present invention, retinal arteries are outlined with a red brush, retinal veins are outlined with a blue brush, and arteriovenous intersections are outlined with a green brush. The deep learning model simultaneously achieves retinal blood vessel segmentation and arteriovenous classification.

[0104] The present invention adopts a 4-layer U-shaped network as our backbone framework, which consists of an encoding path and a decoding path. In the encoding path, each level contains two convolutional layers with a core size of 3, followed by a 2×2 max pooling layer. In the decoding path, each level contains an upsampling layer to restore the size of the feature map. The skip connection between the encoder and decoder integrates shallow low-level texture features and high-level semantic features, thereby retaining more spatial information for better segmentation and classification.

[0105] In addition, the present invention guides the model to focus on the smoothness of the segmentation results through a boundary enhancement module, which consists of multiple pixel-level edge attention maps generated by the Sobel detector.

[0106] For the network that segmented retinal vessels, the output image contained four categories: arteries, veins, bifurcations, and background.

[0107] In a specific embodiment, optic disc segmentation: Unlike the complex topological structure of retinal blood vessels, the optic disc is an elliptical segmentation target. Therefore, the present invention realizes optic disc segmentation in an unlabeled manner by transferring knowledge from the public fundus dataset IDRiD. The present invention trains the model by fusing the data of Shougang Research and IDRiD data with different weights. During the training process, the model initially gives a higher weight to the IDRiD data and gradually increases the weight of the unlabeled data. Therefore, it gradually extracts more effective image features to segment the optic disc. The network for segmenting the optic disc is no different in structure from the network for segmenting the retinal blood vessels, and both use the same U-shaped network.

[0108] However, due to the different specific tasks, the two networks have different outputs. For the network for segmenting the optic disc, the output image contains the optic disc category and the background category.

[0109] S103: performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0110] In a specific embodiment, after the optic disc is identified, the model automatically determines the optic disc diameter and the optic disc center. In order to measure the potential correlation between the curvature of retinal blood vessels and MACE, the present invention first performs binarization processing on the obtained retinal image, and then performs skeletonization processing to obtain a skeleton image. At the same time, the present invention extracts the center line of the binarized retinal image and obtains the distance from each pixel on the blood vessel to the center line; then the present invention retains the blood vessels within the range of 1-1.5 times the optic disc diameter of the optic disc center, that is, the blood vessels in the target ring area.

[0111] S104: Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0112] In one embodiment, the screening is to identify the blood vessel bifurcation point through the graph data structure, and then screen out the blood vessels containing the blood vessel bifurcation point to obtain the screened image;

[0113] Preferably, the graph data structure adopts an undirected graph for recognition, and the recognition constructs a graph node for each pixel point of the skeleton image, and then constructs an edge for adjacent pixel points to form an undirected graph; the degree of the node of the undirected graph is judged, and when the degree of the node of the undirected graph is greater than a preset threshold, it is determined to be a bifurcation point.

[0114] In one embodiment, the screening further includes screening the too short or too thin blood vessels in the skeleton image, obtaining the primary screening blood vessels after screening the too short or too thin blood vessels in the skeleton image, then identifying the blood vessel bifurcation points in the primary screening blood vessels, and screening the blood vessels containing the blood vessel bifurcation points to obtain the screened image;

[0115] Preferably, the specific steps of screening out the too short or too thin blood vessels are:

[0116] The first step is to binarize the skeleton image to obtain a binary image;

[0117] The second step is to extract the centerline of the binary image to obtain the centerline, and obtain the distance from each pixel on the blood vessel to the centerline based on the centerline;

[0118] The third step is to calculate the average value of the distances to obtain an average distance and use the average distance as the blood vessel width;

[0119] Step 4: Delete the blood vessels whose width is less than a preset threshold, and delete the blood vessels whose length is less than m pixels to obtain the initial screening blood vessels, where m is a positive number less than or equal to 20;

[0120] Optionally, the preset threshold is to sort all the calculated blood vessel widths, select the first n blood vessel widths and calculate the average to obtain a width mean, where n is a natural number greater than 2; and use the value obtained by subtracting k times the standard deviation from the width mean as the threshold, where k is a natural number greater than 1;

[0121] Preferably, the step of screening out blood vessels that are too short or too thin also includes selecting a target area, after obtaining the distance from each pixel on the blood vessel to the center line, taking the center of the optic disc as the center and an area formed by c times the diameter of the optic disc as the target area, and calculating the blood vessels in the target area to obtain the blood vessel width, where c is a natural number greater than or equal to 1.

[0122] In a specific embodiment, for the blood vessels in the target area, the present invention will delete the blood vessels that are too short and too thin. The present invention uses the average distance from all pixels on the blood vessel to the center line as the blood vessel width, and uses the mean of the width of the six thickest blood vessels minus 2.5 times the standard deviation as the threshold. The short blood vessels with a width less than the threshold are deleted, and the short blood vessels with a length less than 20 pixels are deleted. In this way, the present invention completes the post-processing work of arteriovenous classification.

[0123] In a specific embodiment, identification of bifurcated blood vessels: for the skeletonized image after post-processing, we use an undirected graph to store its pixels, construct a graph node for each pixel, and construct an edge for each pair of adjacent pixels. The present invention considers nodes with a degree greater than 2 as bifurcations, and deletes all blood vessels with bifurcations.

[0124] S105: Calculate the curvature of the blood vessel based on the screened image.

[0125] In one embodiment, the calculation is performed by fitting the coordinates of the pixel points in the filtered image into a curve to obtain a blood vessel approximate curve, and then performing curvature calculation based on the curve to obtain the curvature of the blood vessel;

[0126] Preferably, the pixel point coordinates are fitted as a curve by fitting a cubic polynomial;

[0127] Preferably, the pixel coordinates in the screened image are fitted into a curve and replaced by: connecting two ends of the blood vessel in the screened image to obtain a connecting line midpoint and using it as a rotation center, rotating the blood vessel based on the rotation center so that the connecting line midpoint is parallel to the horizontal axis in the horizontal axis coordinate to obtain the rotated blood vessel, mapping the rotated blood vessel to the horizontal and vertical coordinates to obtain updated pixel coordinates, and fitting the updated pixel coordinates into a curve to obtain a blood vessel approximate curve;

[0128] Optionally, the curve further includes one or more of the following calculations: arc length calculation of the curve, integral calculation of the square of the curvature of the curve, cTORT calculation;

[0129] Optionally, the calculation formula of cTORT is:

[0130]

[0131] Where tsc(C) represents the integral of the square of the curvature, s(C) represents the arc length of the curve, and C represents the curve.

[0132] In a specific embodiment, pixel rotation and curve fitting: a series of horizontal and vertical coordinates of pixel points can be used to describe a blood vessel, but in the original image, the horizontal and vertical coordinates may not form a one-to-one mapping function relationship, and this method is necessary to obtain the fitted analytical expression of the pixel points when solving the curvature. Therefore, the present invention will take the midpoint of the line connecting the two ends of the blood vessel as the rotation center, rotate the blood vessel so that the midpoint of the line connecting the two ends of the blood vessel is parallel to the horizontal axis, thereby obtaining a series of new horizontal and vertical coordinates of pixel points, and perform cubic polynomial curve fitting on these new pixel coordinates to obtain the blood vessel approximate curve C.

[0133] In a specific embodiment, assuming that each blood vessel can be described by a curve C=(x(t), y(t)) defined on the interval [t0, t1], then:

[0134] The arc length formula of curve C is:

[0135]

[0136] The curvature formula of curve C at t is:

[0137]

[0138] The integral formula of the square of the curvature of curve C is:

[0139]

[0140] cTORT formula:

[0141]

[0142] Finally, the present invention uses the above formula to perform numerical calculations to obtain the curvature of the blood vessel.

[0143] In one specific embodiment, 149 single fundus images were randomly selected and manually delineated from all 3791 subjects, of which 120 images were randomly selected for model training and the remaining 29 images were used for internal model validation. This requires not only distinguishing the correct vessel type (artery or vein), but also accurately outlining the boundaries of the vessels on the fundus photos.

[0144] All images of the 792 participants were then identified and measured by the deep learning model. Experienced ophthalmologists checked the error rate of the deep learning model in identifying retinal blood vessels using all cases. The error rate for artery identification was 1.26%, the error rate for vein identification was 0.79%, and the overall error rate was 1.03%.

[0145] In a specific embodiment, the present invention conducts an experiment on the relationship between the curvature of retinal vessels measured by deep learning and new MACE, and the experimental results are as follows Figure 4 The Kaplan-Meier curve of the risk of MACE in the cTORTa dichotomy is shown. Figure 5 The Kaplan-Meier curve of the risk of MACE in the cTORTv dichotomy shown in the figure indicates that the larger the cTORT, the higher the cumulative risk of MACE.

[0146] In addition, multivariate Cox regression analysis showed that the risk of new MACE in the cTORTa larger group (B2: 1.85E-4 ~ 36.08E-4) increased by 52% (HR = 1.52, 95% CI: 1.11-2.08, P = 0.009) compared with the cTORTa smaller group (B1: 0.06E-4 ~ 1E-4).

[0147] In a specific embodiment, the present invention calculates the curvature tortuosity arteriole (cTORTa) and the curvature tortuosity venule (cTORTv), which are important for understanding the association between retinal vascular curvature and MACE.

[0148] The experimental results show that this method has an extremely low error rate and shows a high degree of accuracy, proving its potential for application in fundus image analysis and cardiovascular health risk assessment. Overall, this patented technical solution, by combining advanced image preprocessing methods and deep learning models, not only improves the quality and consistency of data obtained from automated fundus image analysis, but also has important clinical value in MACE risk assessment.

[0149] In a specific embodiment, retinal microvascular abnormalities are associated with hypertension, arteriosclerosis, and cardiovascular disease. In addition, retinal artery stenosis may reflect the severity of cardiovascular disease complications and serve as an independent risk factor for cardiovascular disease. Therefore, identifying retinal vascular abnormalities and quantifying vascular diameters may play an important role in long-term monitoring of cardiovascular disease. The development of digital retinal photography and computer technology has enabled the practical application of the Parr-Hubbard formula. Integrative Vessel Analysis (IVAN) is a computer-assisted software that can identify and quantitatively evaluate retinal arteries and veins. It has been widely used in several population-based studies over the past 10 years with high repeatability and reliability. In addition, changes in retinal vascular diameter are associated with hypertension, coronary artery disease, and stroke. However, IVAN is a user-dependent and time-consuming software that requires manual annotation and correction by doctors. Therefore, the present invention also uses deep learning to calculate the diameter of retinal vessels, while comparing the measurement results of IVAN software.

[0150] In a specific embodiment, a nested case-control study was established including 155 participants with new carotid plaques during the 6-year follow-up and 155 control participants matched by the same sex and age ±3 years. Then, CRAE, CRVE, and AVR were measured using IVAN software. To study the association between deep learning and IVAN software-measured retinal vessel diameter and hypertension, 310 participants in this nested control study, including 103 baseline hypertensives and 207 baseline non-hypertensives, were selected because they had both CRAE, CRVE, and AVR data measured by IVAN software and data measured by the established deep learning model. Hypertension was defined as systolic blood pressure ≥140 mmHg, diastolic blood pressure ≥90 mmHg, self-reported history of hypertension, or taking antihypertensive medication.

[0151] Color fundus photographs were taken of each eye using a CR6-45NM fundus camera (Canon Inc., Tokyo, Japan), with the images centered on the optic disc (diabetic retinopathy research standard zone 1) and the macula (diabetic retinopathy research standard zone 2). To assess changes in retinal vessel diameter, retinal vessel diameters were measured using computer-assisted quantitative assessment software (IVAN; University of Wisconsin-Madison) as a comparison result for the present method. For each photograph, retinal vessels were analyzed in an area 0.5-1.0 optic disc diameters from the optic disc edge. After manually selecting six or more of the largest arteries and veins, the IVAN software automatically summarized the vessel diameters into a pair of indices. Retinal vessel diameters were normalized to central retinal artery equivalents (CRAE) and central retinal vein equivalents (CRVE) using the modified Knudtson-Parr-Hubbard formula. AVR was calculated as the ratio of CRAE to CRVE.

[0152] A deep learning model based on a U-shaped network structure is used for accurate segmentation of retinal vessels, classification of arteries and veins, and automatic segmentation of the optic disc. After the artery and vein classification is completed, the target area is first determined based on the optic disc, and then the segmented vessels are processed by distance transformation, extraction of vessel centerlines, vessel screening, and determination of vessel bifurcation points. These steps are helpful for subsequent vessel diameter measurement and analysis. Finally, the central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), and artery-vein ratio (AVR) are calculated.

[0153] In one embodiment, the method implements the method for calculating the curvature of the retinal blood vessels, and the step of calculating the curvature of the blood vessels based on the screened image is replaced by: calculating the diameter of the blood vessels based on the screened image;

[0154] Optionally, the calculation of the diameter includes one or more of the following parameters: vessel diameter, average vessel thickness, CRAE, CRVE, AVR;

[0155] Preferably, the step of identifying the blood vessel bifurcation point is replaced by: first extracting the center line of the skeleton image and obtaining the topological structure of blood vessel distribution based on the center line, and then identifying the bifurcation point in the topological structure by a sliding window method;

[0156] Preferably, the step of removing the blood vessels containing the blood vessel bifurcation point is replaced by: taking the blood vessels near the center of the optic disc in the target area as the main blood vessels, wherein the blood vessels containing the blood vessel bifurcation point in the main blood vessels are removed, and the bifurcated blood vessels are removed and the main trunk is retained; the blood vessels whose bifurcation points are less than d pixels away from the boundary of the target area are removed from the main trunk and the bifurcated blood vessels are retained, where d is a natural number greater than 1; and finally, the image after the removal is obtained;

[0157] Preferably, the average blood vessel thickness is calculated by first calculating the blood vessel thickness and then averaging the blood vessel thickness to obtain the average blood vessel thickness;

[0158] Preferably, the blood vessel thickness is the distance between the center line and the blood vessel boundary;

[0159] Preferably, the blood vessel diameters in the screened images are sorted, and the first h blood vessel diameters are selected and averaged to obtain the average blood vessel diameter, where h is a natural number greater than 1.

[0160] In one embodiment, the method for calculating the retinal diameter includes:

[0161] Acquire fundus images;

[0162] Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels;

[0163] Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0164] Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0165] The screening includes screening out blood vessels that are too short or too thin on the skeleton image, obtaining primary screening blood vessels after screening out blood vessels that are too short or too thin on the skeleton image, and then identifying blood vessel bifurcation points on the primary screening blood vessels, and screening out blood vessels containing blood vessel bifurcation points to obtain a screened image; wherein the blood vessel bifurcation points are identified through a graph data structure, and then the blood vessels containing the blood vessel bifurcation points are screened out to obtain a screened image;

[0166] The diameter of the blood vessel is calculated based on the screened image.

[0167] In one embodiment, the method for calculating the retinal diameter includes:

[0168] Acquire fundus images;

[0169] Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels;

[0170] Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0171] Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0172] The screening includes screening the too short or too thin blood vessels of the skeleton image, obtaining the primary screening blood vessels after screening the too short or too thin blood vessels of the skeleton image, then identifying the blood vessel bifurcation points of the primary screening blood vessels, and screening the blood vessels containing the blood vessel bifurcation points to obtain the screened image; wherein, firstly, the center line of the skeleton image is extracted and the topological structure of blood vessel distribution is obtained based on the center line, and then the bifurcation points in the topological structure are identified by a sliding window method; and then the blood vessels containing the blood vessel bifurcation points are screened to obtain the screened image;

[0173] The diameter of the blood vessel is calculated based on the screened image.

[0174] In one embodiment, the method for calculating the retinal diameter includes:

[0175] Acquire fundus images;

[0176] Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels;

[0177] Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0178] Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0179] The screening includes screening the too short or too thin blood vessels of the skeleton image, obtaining the primary screening blood vessels after screening the too short or too thin blood vessels of the skeleton image, and then identifying the blood vessel bifurcation points of the primary screening blood vessels, and screening the blood vessels containing the blood vessel bifurcation points to obtain the screened image; wherein, firstly extract the center line of the skeleton image and obtain the topological structure of blood vessel distribution based on the center line, and then identify the bifurcation points in the topological structure by a sliding window method; the blood vessels close to the center of the optic disc in the target area are taken as the main blood vessels, wherein the bifurcated blood vessels containing the blood vessel bifurcation points in the main blood vessels are screened out and the bifurcated blood vessels are screened out and the main trunks are retained; the blood vessels whose bifurcation points are less than d pixels away from the boundary of the target area are screened out and the bifurcated blood vessels are retained, where d is a natural number greater than 1; and finally the screened image is obtained;

[0180] The diameter of the blood vessel is calculated based on the screened image.

[0181] In one embodiment, the method for calculating the retinal diameter includes:

[0182] Acquire fundus images;

[0183] Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels;

[0184] Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image;

[0185] Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image;

[0186] The screening includes screening out the too short or too thin blood vessels of the skeleton image, obtaining the primary screening blood vessels after screening out the too short or too thin blood vessels of the skeleton image, and then identifying the blood vessel bifurcation points of the primary screening blood vessels, and screening out the blood vessels containing the blood vessel bifurcation points to obtain the screened image; wherein the blood vessel bifurcation points are identified through a graph data structure; the blood vessels near the center of the optic disc in the target area are taken as the main blood vessels, wherein the bifurcated blood vessels containing the blood vessel bifurcation points in the main blood vessels are screened out and the bifurcated blood vessels are screened out and the main trunks are retained; the blood vessels whose bifurcation points are less than d pixels away from the boundary of the target area are screened out and the bifurcated blood vessels are retained, where d is a natural number greater than 1; and finally the screened image is obtained;

[0187] The diameter of the blood vessel is calculated based on the screened image.

[0188] In one specific embodiment, a distance transform is applied to the binary fundus vessel image, which assigns a value to each foreground pixel (vessel) in the binary image, which is its distance to the nearest background pixel (non-vessel, vessel boundary).

[0189] Using the Skeleton method, all foreground (vessel) pixels of the binary image are thinned to a center line with a pixel width to obtain a skeletonized image.

[0190] On the skeletonized image, the diameter of each centerline pixel can be determined by looking at the value of the corresponding position in the original distance transform image. Since each pixel value in the distance transform image represents the distance to the nearest background pixel, that is, the distance to the edge of the vessel, this value is actually the radius. Therefore, the diameter is twice this distance value.

[0191] In a specific embodiment, the present invention focuses on the main retinal blood vessel trunks in the target ring area near the optic disc. In most cases, the present invention selects the blood vessel trunks to calculate the average blood vessel diameter. However, when the bifurcation point is very close to the inner boundary of the target ring area (less than 5 pixels in the present invention), the local swelling at the bifurcation point may greatly affect the measurement of the blood vessel trunk. Therefore, in this case, the present invention abandons the trunk and calculates the average blood vessel diameter based on the branches. Finally, relatively shorter segmented blood vessel segments with a centerline length of less than 20 pixels are deleted to avoid instability. Considering that doctors using IVAN software only consider the main blood vessels around the optic disc, all small blood vessels (2.5 standard deviations thinner than the diameter mean) are deleted from the final arteriovenous classification results.

[0192] In a specific embodiment, the present invention selects the main blood vessel to calculate the average blood vessel diameter. Therefore, it is necessary to process the blood vessels with bifurcated blood vessels, use a sliding window method to identify the bifurcation point of the blood vessel, discard the bifurcated blood vessels and retain the main blood vessel.

[0193] In a specific embodiment, since each fundus image contains about 6 blood vessels, each fundus image will obtain about 6 blood vessel diameter values ​​after calculating the blood vessel diameter. In order to comprehensively represent the diameters of all blood vessels, consider taking the average value of all blood vessel diameter values ​​as the representative of the blood vessel diameter of the fundus image.

[0194] In a specific embodiment, in order to measure the potential correlation between retinal blood vessel diameter and cardiovascular disease, the present invention includes blood vessels within the range of 1-1.5 times the diameter of the optic disc center, that is, blood vessels within the target ring area. A distance transform is performed on the segmentation result, and the centerline (skeleton) of the blood vessel is extracted. The topological structure is analyzed based on the centerline of the blood vessel, and then the blood vessel segments that are too short and too thin are deleted and the bifurcation points of the segmented blood vessels are determined. The blood vessels are segmented at the bifurcation points, and the blood vessel segments close to the center of the optic disc are defined as the blood vessel trunk, otherwise they are defined as blood vessel branches. In order to calculate the average blood vessel thickness in the target ring area, the blood vessel trunk is selected in most cases, but when the distance between the bifurcation point and the boundary of the target ring area is less than 5 pixels, the branch is used instead of the trunk. Among them, the thickness of the blood vessel is defined as the distance between the centerline and the blood vessel boundary.

[0195] Finally, three key parameters including CRAE surrogate, CRVE surrogate and AVR were calculated according to the modified Parr-Hubbard formula within the target annulus area.

[0196] In a specific embodiment, the present invention tests the model proposed by the present invention on independent data of 482 Tongren studies and 310 Shougang studies. The correlation coefficients of CRAE, CRVE and AVR between the automatic output results of deep learning and the manual determination results of IVAN software were 0.38, 0.55 and 0.41 (all P<0.001) for Tongren studies and 0.56, 0.52 and 0.51 (all P<0.001) for Shougang studies. The correlation coefficients of the difference between the retinal vascular diameter measured by IVAN and deep learning and the average value were 0.02 (P=0.704) and -0.09 (P=0.110) in Tongren studies and Shougang studies, respectively.

[0197] The experimental results show that compared with the traditional semi-automatic software IVAN, the deep learning model of the present invention performs well in terms of efficiency and diagnostic value, proving its potential for application in fundus image analysis and cardiovascular health risk assessment. Overall, this technical solution, by combining advanced image preprocessing methods and deep learning models, not only improves the quality and consistency of data obtained from automated fundus image analysis, but also has important clinical value in hypertension risk assessment.

[0198] In one embodiment, the disease risk is predicted based on the retinal blood vessels, the retinal blood vessel curvature is calculated by executing the method for calculating the retinal blood vessel curvature; and the retinal blood vessel diameter is calculated by executing the method for calculating the retinal blood vessel diameter;

[0199] The curvature of the retinal blood vessels and the diameter of the retinal blood vessels are input into a prediction model to obtain a disease prediction result.

[0200] In one embodiment, the method implements the method for calculating the curvature of the retinal blood vessels, and simultaneously calculates the diameter of the blood vessels based on the screened images, thereby obtaining the curvature of the retinal blood vessels and the diameter of the retinal blood vessels;

[0201] The curvature of the retinal blood vessels and the diameter of the retinal blood vessels are input into a prediction model to obtain a disease prediction result.

[0202] In one embodiment, diseases predicted based on the curvature and diameter of retinal blood vessels include:

[0203] Major adverse cardiovascular events (including cardiovascular death, stroke and myocardial infarction), hypertension.

[0204] Figure 2 A schematic diagram of a device provided by an embodiment of the present invention specifically includes:

[0205] A memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, when the program instructions are executed, any one of the above-mentioned methods for calculating the curvature of the retinal blood vessels or the system executes the method for calculating the diameter of the retinal blood vessels or the method for predicting disease risks based on the retinal blood vessels is executed.

[0206] A system, comprising:

[0207] The system executes the method for calculating the curvature of the retinal blood vessels or the method for calculating the diameter of the retinal blood vessels or the method for predicting disease risks based on the retinal blood vessels.

[0208] A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs any one of the above-mentioned methods for calculating the curvature of a retinal vessel, or performs the method for calculating the diameter of a retinal vessel performed by the system, or performs a method for predicting disease risks based on retinal vessels.

[0209] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of the method relative to the default setting. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units. A person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above-mentioned embodiments may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0210] A person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned medium storage can be a read-only memory, a disk or an optical disk, etc.

[0211] The above is a detailed introduction to a computer device provided by the present invention. For a person skilled in the art, according to the concept of the embodiments of the present invention, there may be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for calculating retinal vascular curvature, characterized in that: The method comprises: Acquire fundus images; Perform blood vessel segmentation on fundus images to obtain arterial and venous vessels; Performing skeleton extraction on the arterial blood vessels and the venous blood vessels to obtain a skeleton image; Screening out blood vessels with bifurcation points in the skeleton image to obtain a screened image; the screening is to identify the blood vessel bifurcation points through a graph data structure, and then screen out blood vessels containing the blood vessel bifurcation points to obtain the screened image; The curvature of the blood vessel is calculated based on the screened image; The calculation is performed by fitting the coordinates of the pixel points in the screened image into a curve to obtain an approximate curve of the blood vessel, and then performing curvature calculation based on the curve to obtain the curvature of the blood vessel.

2. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The graph data structure adopts an undirected graph for recognition. The recognition constructs a graph node for each pixel point of the skeleton image, and then constructs an edge for adjacent pixel points to form an undirected graph; the degree of the node of the undirected graph is judged, and when the degree of the node of the undirected graph is greater than a preset threshold, it is determined to be a bifurcation point.

3. The method for calculating the curvature of retinal blood vessels according to claim 1, characterized in that: The screening also includes screening out blood vessels that are too short or too thin in the skeleton image, obtaining primary screening blood vessels after screening out blood vessels that are too short or too thin in the skeleton image, then identifying blood vessel bifurcation points in the primary screening blood vessels, and screening out blood vessels containing blood vessel bifurcation points to obtain a screened image.

4. The method for calculating the curvature of retinal blood vessels according to claim 3, characterized in that: The specific steps of screening out blood vessels that are too short or too thin are: The first step is to binarize the skeleton image to obtain a binary image; The second step is to extract the centerline of the binary image to obtain the centerline, and obtain the distance from each pixel on the blood vessel to the centerline based on the centerline; The third step is to calculate the average value of the distances to obtain an average distance and use the average distance as the blood vessel width; In the fourth step, the blood vessels whose width is less than a preset threshold are deleted, and the blood vessels whose length is less than m pixels are deleted to obtain the initial screening blood vessels, where m is a positive number less than or equal to 20.

5. The method for calculating the curvature of retinal blood vessels according to claim 4, characterized in that: The preset threshold is to sort all the calculated blood vessel widths, select the first n blood vessel widths and calculate the average to obtain the width mean, where n is a natural number greater than 2; and use the value obtained by subtracting k times the standard deviation from the width mean as the threshold, where k is a natural number greater than 1.

6. The method for calculating the curvature of retinal blood vessels according to claim 4, characterized in that: The step of screening out blood vessels that are too short or too thin also includes selecting a target area, after obtaining the distance from each pixel on the blood vessel to the center line, taking the center of the optic disc as the center of a circle and an area formed by c times the diameter of the optic disc as the target area, and calculating the blood vessel width of the blood vessel in the target area, where c is a natural number greater than or equal to 1.

7. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The segmentation is performed by one or more of the following segmentation models: U-net, FCN, SegNet, DeepLab v3, Faster R-CNN, PSPnet, ESPNet, Mask R-CNN.

8. The method for calculating the curvature of retinal blood vessels according to claim 7, characterized in that: The segmentation is performed through a U-net network framework and a boundary enhancement module, and edge attention mapping features are generated by performing edge detection on the fundus image, and the edge attention mapping features are input into the convolution layer of the U-net network framework encoding module to enhance feature representation.

9. The method for calculating the curvature of retinal blood vessels according to claim 7, characterized in that: The segmentation also includes optic disc segmentation, performing optic disc segmentation on the fundus image to obtain an optic disc; and selecting a target area based on the optic disc.

10. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The method also includes data preprocessing, which includes one or more of the following: dark channel prior assumption, image cropping and filling, grayscale, geometric transformation, binarization, and histogram equalization.

11. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The pixel point coordinates are fitted into a curve by a cubic polynomial.

12. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The pixel coordinates in the screened image are fitted into a curve and replaced by: connecting the two ends of the blood vessel in the screened image to obtain the midpoint of the line and using it as the rotation center, rotating the blood vessel based on the rotation center so that the midpoint of the line is parallel to the horizontal axis in the horizontal axis coordinate to obtain the rotated blood vessel, mapping the rotated blood vessel to the horizontal and vertical coordinates to obtain updated pixel coordinates, and fitting the curve based on the updated pixel coordinates to obtain a blood vessel approximate curve.

13. The method for calculating retinal vessel curvature according to claim 1, characterized in that: The curve also includes one or more of the following calculations: arc length calculation of the curve, integral calculation of the square of the curvature of the curve, and cTORT calculation; The calculation formula of cTORT is: Where tsc(C) represents the integral of the square of the curvature, s(C) represents the arc length of the curve, and C represents the curve.

14. A method for calculating retinal blood vessel diameter, characterized in that: include: The method implements the method for calculating the curvature of the retinal blood vessels as described in any one of claims 1 to 13, and replaces the step of calculating the curvature of the blood vessels based on the screened image with: calculating the diameter of the blood vessels based on the screened image.

15. The method for calculating the diameter of retinal blood vessels according to claim 14, characterized in that: The calculation of the diameter includes one or more of the following parameters: blood vessel diameter, average blood vessel thickness, CRAE, CRVE, and AVR.

16. The method for calculating the retinal blood vessel diameter according to claim 14, characterized in that: The step of identifying the blood vessel bifurcation point is replaced by: first extracting the center line of the skeleton image and obtaining the topological structure of the blood vessel distribution based on the center line, and then identifying the bifurcation point in the topological structure by a sliding window method.

17. The method for calculating the retinal blood vessel diameter according to claim 14, characterized in that: The step of screening out the blood vessels containing the blood vessel bifurcation point is replaced by: taking the blood vessels near the center of the optic disc in the target area as the main blood vessels, wherein the blood vessels containing the blood vessel bifurcation point in the main blood vessels are screened out and the bifurcated blood vessels are screened out and the main trunk is retained; the blood vessels whose bifurcation points are less than d pixels away from the boundary of the target area are screened out and the main trunk is retained, where d is a natural number greater than 1; and finally, the screened image is obtained.

18. The method for calculating the diameter of retinal blood vessels according to claim 15, characterized in that: The average blood vessel thickness is calculated by first calculating the blood vessel thickness and then averaging the blood vessel thickness to obtain the average blood vessel thickness.

19. The method for calculating the diameter of retinal blood vessels according to claim 18, characterized in that: The blood vessel thickness is the distance between the center line and the blood vessel boundary.

20. The method for calculating retinal blood vessel diameter according to claim 14, characterized in that: The vessel diameters in the screened images are sorted, and the first h vessel diameters are selected and averaged to obtain the average vessel diameter, where h is a natural number greater than 1.

21. A method for predicting disease risk based on retinal blood vessels, characterized in that: include: The method implements the method for calculating the curvature of the retinal blood vessels according to any one of claims 1 to 13 to obtain the curvature of the retinal blood vessels; The method implements the method for calculating the retinal vessel diameter according to any one of claims 14 to 20 to obtain the retinal vessel diameter; The curvature of the retinal blood vessels and the diameter of the retinal blood vessels are input into a prediction model to obtain a disease prediction result.

22. A device for calculating retinal blood vessel curvature, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions, and the processor is used to call program instructions, and when the program instructions are executed, the method for calculating the curvature of the retinal blood vessels described in any one of claims 1-13 is implemented.

23. A device for calculating retinal blood vessel diameter, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions, and the processor is used to call program instructions, and when the program instructions are executed, the method for calculating the retinal blood vessel diameter according to any one of claims 14 to 20 is implemented.

24. A device for predicting disease risk based on retinal blood vessels, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions, and the processor is used to call program instructions, and when the program instructions are executed, the method for predicting disease risk based on retinal blood vessels as described in claim 21 is implemented.

25. A computer-readable storage medium, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, the method for calculating the curvature of the retinal blood vessels described in any one of claims 1 to 13 is implemented, or the method for calculating the diameter of the retinal blood vessels described in any one of claims 14 to 20 is implemented, or the method for predicting disease risks based on retinal blood vessels described in claim 21 is implemented.

Images