Method and device for segmenting retinal blood vessels in fundus image

Abstract

The invention belongs to the technical field of computer vision and medical image processing, particularly relates to a method and a device for segmenting retinal blood vessels in fundus images, and aims to solve the problems of poor adaptability to different image quality, weak anti-interference capability to compound noise and low blood vessel reconstruction accuracy in the prior art. The method comprises the following steps: acquiring a to-be-processed eye fundus image; analyzing the eye fundus image to determine a prior analysis index and noise classification information, and further determining a processing parameter; processing the eye fundus image by adopting the processing parameters and executing noise elimination processing to generate an initial blood vessel segmentation image; identifying break points and candidate target points in the initial blood vessel segmentation map; and performing multi-dimensional verification between the interruption point and each candidate target point to determine optimal connection and generate a reconstructed retinal blood vessel segmentation map. According to the method, through adaptive parameter adjustment and multi-dimensional verification reconstruction, the accuracy and robustness of segmentation and the adaptability to complex lesion images are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision and medical image processing, and particularly relates to a retinal blood vessel segmentation method and device in fundus images. BACKGROUND

[0002] The fundus retinal blood vessels are the only blood vessel system that can be directly and non-invasively observed in the human body. Changes in the morphological structure, such as the tortuosity, diameter, branching angle, arteriovenous diameter ratio (AVR) of the blood vessels, are closely related to many systemic cardiovascular and cerebrovascular diseases such as hypertension, diabetes, and arteriosclerosis. Therefore, the segmentation of retinal blood vessels in fundus images is one of the key technologies for computer-aided diagnosis, and its accuracy directly affects the early screening and condition assessment of various cardiovascular and cerebrovascular diseases.

[0003] Existing retinal blood vessel segmentation methods, such as the Hessian matrix method based on multi-scale analysis, can effectively enhance the linear structure of blood vessels, but usually use fixed calculation parameters. This "one-size-fits-all" approach has poor adaptability when facing fundus images of different sources and qualities. For example, for low-contrast or high-noise images, fixed parameters can easily lead to missed detection of a large number of fine blood vessels or false detection of noise; and for complex images with dense blood vessels or lesions, the segmentation accuracy is also difficult to guarantee.

[0004] In addition, after preliminary segmentation, due to image artifacts, lesion occlusion, or algorithm limitations, the blood vessel network often appears to be broken and discontinuous. Traditional blood vessel reconstruction methods mostly rely on simple shortest path connection or single morphological rules, which can easily produce physiologically unreasonable false connections when dealing with complex topological structures such as blood vessel intersections and branches, and can mistakenly connect blood vessel segments that do not belong to the same branch, which seriously affects the reliability of subsequent blood vessel morphological quantitative analysis. At the same time, the complex noise formed by the high-brightness reflection of the optic disc area, hemorrhagic spots, and exudates, etc. also brings great challenges to accurate segmentation, and the existing technology lacks an effective mechanism for hierarchical and targeted suppression of these different types of noise.

[0005] Therefore, how to design a retinal blood vessel segmentation method that can adaptively adjust parameters to adapt to different image qualities, intelligently verify and accurately reconstruct interrupted blood vessels, and effectively suppress the interference of complex noise, has become a technical problem to be solved in the field. SUMMARY

[0006] In order to solve the above problems in the prior art, i.e. to solve the problems of poor adaptability of the segmentation algorithm to different image qualities, weak anti-interference ability to complex noise, and low accuracy of blood vessel reconstruction in the prior art, on the one hand, the present application provides a retinal blood vessel segmentation method in a fundus image, comprising the following steps:

[0007] obtaining an eye fundus image to be processed;

[0008] analyzing the eye fundus image to determine at least one priori analysis index representing distribution characteristics of the image and to identify noise regions in the image to generate noise classification information;

[0009] dynamically determining processing parameters for extracting blood vessel segmentation features based on the priori analysis index and the noise classification information;

[0010] processing the eye fundus image using the dynamically determined processing parameters and performing noise removal processing based on the noise classification information to generate an initial blood vessel segmentation map containing interrupted blood vessels;

[0011] identifying interruption points of blood vessels in the initial blood vessel segmentation map and candidate destination points on a blood vessel skeleton;

[0012] performing multi-dimensional verification between the interruption points and each of the candidate destination points to determine a destination point corresponding to the interruption point and to connect the destination point to the interruption point to generate a reconstructed retinal blood vessel segmentation map.

[0013] In some preferred embodiments, the priori analysis index includes at least one of the following:

[0014] a blood vessel density obtained by statistical analysis of a sliding window on the eye fundus image;

[0015] a noise intensity calculated by a ratio of a global gray variance and an average edge gradient of the eye fundus image;

[0016] a gray uniformity calculated by a skewness of a gray distribution of the eye fundus image.

[0017] In some preferred embodiments, the processing parameters include at least a window size for Hessian matrix calculation and a feature screening threshold for screening blood vessel features; and the dynamically determining the processing parameters for extracting blood vessel segmentation includes:

[0018] matching an initial window size according to the blood vessel density and the noise intensity;

[0019] matching an initial feature screening threshold according to the noise intensity and the gray uniformity;

[0020] respectively adaptively adjusting the initial window size and the initial feature screening threshold for different types of noise regions according to the noise classification information.

[0021] In some preferred embodiments, the generating the initial blood vessel segmentation map containing interrupted blood vessels includes:

[0022] perform adaptive Hessian matrix feature extraction on the fundus image based on the dynamically determined processing parameter, to generate a blood vessel feature map;

[0023] perform binarization on the blood vessel feature map, to generate a binary blood vessel map;

[0024] perform type-specific elimination on false blood vessel targets corresponding to noise regions in the binary blood vessel map based on the noise classification information, to generate the initial blood vessel segmentation map.

[0025] In some preferred embodiments, the multi-dimension verification includes:

[0026] topology verification based on blood vessel orientation, diameter, or branching angle;

[0027] texture feature verification based on pixel gray scale, edge, or gradient features on the connection path;

[0028] special node-specific rule adaptation verification for blood vessel intersection points, branching points, and terminal capillary.

[0029] In some preferred embodiments, the determination of the destination point corresponding to the interruption point and the connection include:

[0030] for each candidate connection path connecting the interruption point and each candidate destination point, calculate a comprehensive weighted score thereof, which is obtained by weighted summation of path length of the candidate connection path and matching degree obtained in the topology verification and the texture feature verification;

[0031] determine the candidate connection path with the optimal comprehensive weighted score as an effective blood vessel connection path, and determine the endpoint of the effective blood vessel connection path as the destination point corresponding to the interruption point;

[0032] perform smooth fitting connection along the effective blood vessel connection path using a preset curve fitting algorithm.

[0033] In another aspect, the embodiments of the present application provide a device for segmenting retinal blood vessels in a fundus image, the device comprising:

[0034] an analysis module configured to analyze the acquired fundus image to determine at least one prior analysis indicator representing distribution characteristics of the image, and to identify noise regions in the image to generate noise classification information;

[0035] an image processing module configured to dynamically determine processing parameters based on the prior analysis indicator and the noise classification information, and further configured to process the fundus image using the processing parameters and perform noise elimination processing based on the noise classification information, to generate an initial blood vessel segmentation map containing interrupted blood vessels;

[0036] The intelligent reconstruction module is configured to identify an interruption point of the blood vessels in the initial blood vessel segmentation map and a candidate destination point on a blood vessel skeleton, and perform multi-dimensional verification between the interruption point and each candidate destination point to determine a destination point corresponding to the interruption point and connect the destination point to the interruption point to generate a reconstructed retinal blood vessel segmentation map.

[0037] Advantages of the present application:

[0038] 1) The present application dynamically matches the optimal processing parameters through image prior analysis and noise classification, so that the segmentation algorithm can adapt to different quality and different lesion types of fundus images, can effectively deal with fundus images of different quality, and improves the accuracy and robustness of segmentation;

[0039] 2) A layered anti-interference system is constructed, complex noise such as optic disc and hemorrhage points is identified by K-means clustering, and targeted inhibition and removal is performed at different stages, which significantly reduces the interference of noise on blood vessel segmentation and improves the purity and accuracy of the segmentation result;

[0040] 3) The multi-dimensional verification reconstruction algorithm replaces the simple shortest path connection, accurately judges the connection relationship of the interrupted blood vessels, effectively avoids the wrong connection in the complex areas such as blood vessel intersection and branch, ensures the physiological rationality and topological structure integrity of the reconstructed blood vessel network, and thus guarantees the reliability of the final segmentation result. BRIEF DESCRIPTION OF DRAWINGS

[0041] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings:

[0042] Figure 1 is a flowchart of a method for segmenting retinal blood vessels in a fundus image according to an embodiment of the present application;

[0043] Figure 2 is a logic diagram of adaptive parameter determination according to an embodiment of the present application;

[0044] Figure 3 is a schematic diagram of a blood vessel reconstruction process according to an embodiment of the present application, which clearly shows that the candidate destination point is a data stream, and gradually passes through the multi-dimensional verification process of topological structure verification, texture feature verification and special node rule adaptation, and the interruption point is a fixed reference information participating in the progressive screening logic of each level verification;

[0045] Figure 4 is a structural block diagram of a device for segmenting retinal blood vessels in a fundus image according to an embodiment of the present application;

[0046] Figure 5Fig. 1 is a structural schematic diagram of a computer system of a server for implementing the server of the method, system and electronic device embodiments of the present application. DETAILED DESCRIPTION

[0047] The present application will be further described below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are intended to explain the present application and are not intended to limit the present application. In addition, it should be noted that only the parts related to the present application are shown in the accompanying drawings for the convenience of description.

[0048] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0049] The present application proposes a retinal blood vessel segmentation method and device in fundus images, aiming to solve the problems of poor adaptability of existing segmentation algorithms to different image qualities, weak anti-interference ability to composite noise, and low blood vessel reconstruction accuracy. The present application extracts three core indexes of blood vessel density, noise intensity, and gray uniformity through image prior analysis, dynamically matches the calculation window size (W) of the Heaviside matrix and the feature screening threshold (T), solves the adaptability defects of fixed parameters to different images, and improves the reconstruction accuracy and anti-interference ability.

[0050] In order to more clearly describe the retinal blood vessel segmentation method in fundus images of the present application, the following will be described in conjunction with Figure 1 The method in the embodiments of the present application will be described in detail.

[0051] The retinal blood vessel segmentation method in fundus images of the first embodiment of the present application, as shown in Figure 1 The steps S100-S600 are described in detail as follows:

[0052] S100, acquiring a fundus image to be processed and performing preprocessing.

[0053] The original fundus image to be subjected to blood vessel segmentation is acquired. Since the contrast of the retinal blood vessels in the green channel is the highest in the fundus color image, the green channel image needs to be extracted first.

[0054] Preferably, in order to suppress noise and enhance the contrast between blood vessels and background, the green channel image is preprocessed. The preprocessing includes:

[0055] The mean filtering algorithm is used to filter the green channel image to reduce Gaussian noise and salt and pepper noise in the image;

[0056] An adaptive histogram equalization algorithm is used to enhance the contrast of the filtered green channel image to obtain a preprocessed fundus image. This effectively enhances the local contrast of the image, making the edges of fine blood vessels clearer, while avoiding the problem of excessive noise amplification.

[0057] After preprocessing, an optimized fundus image I is obtained, with a size of M×N (M is the number of rows and N is the number of columns), which serves as the input basis for all subsequent steps.

[0058] S200: Analyze the fundus image to determine at least one prior analysis index characterizing the distribution characteristics of the image, and identify noise regions in the image to generate noise classification information.

[0059] Preferably, in order to quantify the overall quality and content distribution of an image, the prior analysis metrics include at least one of the following:

[0060] Vascular density (VD) obtained from fundus images based on sliding window statistics;

[0061] The noise intensity (NI) is calculated by the ratio of the global gray variance to the mean edge gradient of the fundus image.

[0062] Gray-level uniformity (GU) is calculated from the skewness of the gray-level distribution in fundus images.

[0063] More preferably, in this embodiment, the VD value reflects the density of blood vessels, and its calculation method is as follows:

[0064] By setting multiple non-overlapping sliding windows on the image (e.g., window size of 50×50 pixels), the proportion of suspected blood vessels in each window determined by preliminary threshold segmentation (such as Otsu's method) is counted, and then the average proportion of all windows is calculated to obtain the blood vessel density.

[0065] Specifically, in this embodiment, the vascular density (VD) is:

[0066] ;

[0067] in, Divide the image into a total number of non-overlapping windows. The number of suspected blood vessels in the i-th window (pixels with gray values ​​lower than the background mean are identified as suspected blood vessels by Otsu thresholding). The total number of pixels in a single window; and A higher VD value indicates a denser network of blood vessels; if VD=0, the output will show "no blood vessel image".

[0068] More preferably, in this embodiment, the NI value comprehensively evaluates the noise interference level of the image; by calculating the global grayscale variance of the image and... The NI value is obtained by calculating the ratio of the mean edge gradients of the operator.

[0069] Specifically, in this embodiment, the noise intensity (NI) is:

[0070] ;

[0071] in, The global grayscale variance of the image. for The operator calculates the mean of the marginal gradient; (Avoid denominators of 0); a higher NI value indicates stronger noise, typically... .

[0072] More preferably, in this embodiment, the GU value reflects the uniformity of the image grayscale distribution. The GU value is obtained by calculating the skewness of the image grayscale distribution and performing normalization.

[0073] Specifically, in this embodiment, the gray-scale uniformity (GU) is:

[0074] ;

[0075] ;

[0076] in, The grayscale distribution skewness describes the degree of asymmetry in the grayscale distribution; this embodiment preferably uses... (Based on the maximum skewness of over 5000 clinical samples, covering 20 hospitals and 10 common types of fundus lesions); M is the number of rows, N is the number of columns. For pixels Grayscale value The global grayscale mean. The GU value represents the global grayscale variance of the image. The smaller the absolute value of the GU value's skewness, the closer it is to 1, indicating a more symmetrical and uniform grayscale distribution, generally indicating better image quality. hour, It was determined that the grayscale distribution was extremely uneven.

[0077] In a preferred embodiment, in order to distinguish different types of noise and perform targeted processing, this embodiment uses the K-means clustering algorithm to identify noise regions in the image to generate noise classification information. Specifically, based on the gray value of the pixel, the circularity of the connected region and its area characteristics, the clustering algorithm is used to classify the pixel into one of the following: background / blood vessel, optic disc highlight region, bleeding point region or exudate region.

[0078] For example, the bright areas of the optic disc usually have high grayness and high roundness; hemorrhages have medium to high grayness and relatively high roundness; while exudates are characterized by medium grayness but irregular shape.

[0079] More preferably, in one specific embodiment, identifying noisy regions in an image to generate noise classification information includes:

[0080] Extract a feature vector for each pixel in the image, for example, X = (gray value, circularity, area of ​​the connected region);

[0081] Set the number of clusters K=3 (corresponding to the three main types of noise: optic disc highlights, hemorrhages, and exudates), and use initial cluster centers based on clinical experience to cluster all pixels;

[0082] After clustering is completed, a noise classification image N is generated, in which each pixel is labeled as: 1 (visual disc highlight), 2 (bleeding point) or 3 (exudate); non-noise pixels (background / blood vessels) are uniformly labeled as 0.

[0083] S300. Based on the prior analysis indicators and noise classification information, dynamically determine the processing parameters used to extract blood vessel segmentation features.

[0084] Preferably, the processing parameters include at least the window size for calculating the Hessian matrix and the feature screening threshold for screening vascular features.

[0085] More preferably, the dynamic determination of processing parameters for blood vessel segmentation extraction is based on, for example, Figure 2 As shown, specifically:

[0086] Based on prior analysis metrics and noise classification information, the core vascular feature extraction algorithm—Hessian matrix analysis—dynamically matches optimal processing parameters, including:

[0087] The initial window size (W) is matched based on the blood vessel density and the noise intensity.

[0088] An initial feature screening threshold (T) is matched based on the noise intensity and the gray-level uniformity.

[0089] Based on the noise classification information, the initial window size and initial feature selection threshold for different types of noise regions are adaptively adjusted to obtain the adjusted window size and feature selection threshold.

[0090] More preferably, in a specific embodiment, the optimal initial window size W is queried and matched from a preset rule table based on the graded combination of blood vessel density (VD) and noise intensity (NI).

[0091] For example, images with dense blood vessels (high VD value) and low noise (low NI value) are matched with small windows (e.g., 3x3) to capture details, while images with sparse blood vessels (low VD value) or high noise (high NI value) are matched with large windows (e.g., 7x7) to enhance noise resistance. The window size is selected from 3×3, 5×5, and 7×7, based on retinal vessel width statistics (1-20 pixels). A 3×3 window is suitable for fine blood vessels, a 7×7 window enhances noise resistance, and a larger window will lead to the loss of blood vessel details. The feature selection threshold is selected from [0.3, 0.6]. Below 0.3, a large amount of noise will be introduced, and above 0.6, fine blood vessels will be missed. This range has been experimentally verified to balance the detection rate and false detection rate.

[0092] Specifically, in this embodiment, the preset rule table is referenced in Table 1 below.

[0093] Table 1

[0094] More preferably, in one specific embodiment, an optimal initial feature screening threshold T is matched based on a graded combination of noise intensity (NI) and gray-level uniformity (GU). For example, images with low noise (low NI value) and good gray-level uniformity (high GU value) are matched with a lower T to detect more blood vessels, while images with high noise and poor gray-level uniformity are matched with a higher T to suppress noise.

[0095] Specifically, in this embodiment, the preset rule table is referenced in Table 2 below.

[0096] Table 2

[0097] Further preferably, in a specific embodiment, based on the initial matching parameters, local fine-tuning is performed according to the noise classification map N generated in step S200. For regions identified as optic discs (N=1), W is further increased and T is increased; for hemorrhage or exudation regions (N=2 or 3), W and T are moderately increased, with the adjustment range smaller than that for optic disc regions. This aims to further suppress noise response within specific noise regions, which has a higher priority than the global initial matching result.

[0098] Specifically, in one embodiment, based on statistical analysis of clinical samples, it is ensured that the vascular continuity and noise false detection rate can quickly converge to a preset range after parameter correction. The local fine-tuning preferably includes the following:

[0099] Visual area (N=1): Based on the initial matching window size W +2 (e.g., adjust from the original 3×3 to 5×5), and based on the feature threshold T +0.1;

[0100] Bleeding / exudation area (N=2 or 3): W +1, T +0.05; Normal area (N=0): Keep the original matching parameters unchanged;

[0101] Constraints: The adjusted W must not exceed the range of {3,5,7}, and T must not exceed the range of [0.3,0.6].

[0102] S400. The fundus image is processed using dynamically determined processing parameters, and noise removal processing is performed based on the noise classification information to generate an initial vascular segmentation map containing interrupted blood vessels.

[0103] Preferably, generating the initial vascular segmentation map containing interrupted blood vessels includes:

[0104] Based on the dynamically determined processing parameters, adaptive Heather matrix feature extraction is performed on the fundus image to generate a vascular feature map;

[0105] The vascular feature map is binarized to generate a binary vascular map; wherein the size of the local entropy calculation window is the same as the size of the window used for calculating the Hesser matrix.

[0106] Based on the noise classification information, type-specific removal is performed on pseudo-vascular targets corresponding to the noise region in the binary vascular map to generate the initial vascular segmentation map.

[0107] In a preferred embodiment, regarding the vascular features Figure Two Before valueization, feedback calibration is also included, specifically including:

[0108] Calculate the vascular continuity in the vascular feature map, which represents the proportion of vascular pixels within an effective connected region, and the noise false detection rate, which represents the proportion of isolated suspected vascular pixels.

[0109] When the continuity of the blood vessels is lower than a preset lower limit, the feature selection threshold is lowered and adaptive Hesser matrix feature extraction is re-executed; when the false detection rate of noise is higher than a preset upper limit, the feature selection threshold is increased and adaptive Hesser matrix feature extraction is re-executed; until the preset quality requirements are met.

[0110] The continuity lower limit and false detection rate upper limit are based on The clinical fundus images were statistically analyzed, and the samples covered 20 hospitals and 10 common types of fundus lesions.

[0111] More preferably, in one specific embodiment, reference is made to... Figure 2As shown, the fundus image is processed using dynamically determined processing parameters, and noise removal is performed based on the noise classification information to generate an initial vascular segmentation map containing interrupted blood vessels. Specifically, this includes the following:

[0112] S401. Adaptive Hesser matrix feature extraction: Based on the adjusted window size W and feature selection threshold T obtained in step S300, which are adapted to pixels or regions, the Hesser matrix is ​​calculated on the preprocessed fundus image.

[0113] Specifically, in this embodiment, the adaptive Hesser matrix feature extraction includes:

[0114] 1) For each pixel, construct its Heather matrix:

[0115] ;

[0116] Based on statistical analysis of blood vessel width in fundus images, this embodiment preferably uses a Gaussian difference operator with a standard deviation of 1.0 for pixels in image I. ,exist Calculate the second-order partial derivatives within the window and construct the Hesser matrix. ;in: , These are the second-order partial derivatives in the x and y directions, respectively. The mixed second-order partial derivatives are calculated approximately using the Gaussian difference operator.

[0117] 2) Solving for eigenvalues and ( ), and calculate the vascular characteristic response values. ):

[0118] ;

[0119] Among them, response value The larger the value, the closer it is to 1, indicating a higher pixel density. The higher the likelihood that it is a blood vessel.

[0120] 3) Filter the response values ​​using a dynamically determined threshold T, retaining pixels with response values ​​greater than the dynamically determined feature filtering threshold T, and setting the rest to zero, to generate a grayscale blood vessel feature map:

[0121] ;

[0122] in, The filtered feature map (grayscale value) (Non-binary image, preserving characteristic response intensity information).

[0123] S402, Feedback Calibration:

[0124] To avoid poor feature map quality due to improper parameter matching, a feedback calibration mechanism is introduced. Feature map calculation. Vascular continuity (VC) and noise false positive rate (FNR);

[0125] VC = Total number of blood vessel pixels in the effective connected region / Total number of suspected blood vessel pixels

[0126] FNR = Number of isolated suspected blood vessel pixels / Total number of suspected blood vessel pixels;

[0127] Among them, the effective connected region refers to the area The connected regions of a pixel are used to exclude isolated noise. The difference between the threshold for determining noise regions and the threshold for determining noise regions stems from the different area characteristics of blood vessels and noise.

[0128] If VC is below the preset lower limit (based on clinical sample statistics, this embodiment preferably uses 0.6), it indicates that a large number of blood vessels may be missed, so T should be appropriately reduced (e.g., And re-extract features; if the FNR is higher than the preset upper limit (based on clinical sample statistics, this embodiment preferably has 0.2), it indicates that too much noise has been introduced, so T should be appropriately increased (e.g. Then extract again; this loop continues until both metrics meet the preset requirements to ensure output. (hereinafter referred to as) (quality)

[0129] S403, After calibration Binarization:

[0130] For the calibrated vascular feature map Perform binarization to generate a binary blood vessel map. This embodiment preferably employs a local entropy threshold segmentation algorithm, where the window size for calculating local entropy is consistent with the window size W of the Heather matrix to ensure the consistency of the adaptive logic.

[0131] S404. Noise Removal: Based on the noise classification map N generated in step S200, the portions of noise regions identified as different types are removed using the corresponding morphological removal rules for the binary blood vessel map. Perform type-specific noise removal.

[0132] For example, in the area marked as the optic disc, morphological opening operations are used to remove pseudo-vessels; in the area marked as hemorrhage points, isolated targets with high roundness and small area are removed; in the area marked as exudate, irregular areas are screened and removed. After this step, a relatively clean initial vascular segmentation map is obtained, but it may still contain vascular ruptures. (Binary image, 1 represents blood vessel pixels, 0 represents background).

[0133] Specifically, in this embodiment, the type-specific removal includes: removing high-brightness areas of the optic disc using grayscale thresholding and morphological opening operations; and / or removing areas of bleeding points or exudates based on circularity and area thresholds.

[0134] The noise processing employed in this application is a composite noise layered suppression process. This includes a preprocessing stage where a clustering algorithm analyzes the grayscale value, circularity, and connected region area of ​​pixels to distinguish different types of noise, such as optic disc highlights, bleed points, and exudates, outputting a noise classification map. In the Hesser matrix stage, based on the noise classification map, the parameters of the Hesser matrix are adjusted specifically for different noise regions to suppress noise features. In the post-processing stage, based on the geometric features and morphological characteristics of different noises, type-specific noise removal is performed on the binary blood vessel map. This achieves targeted suppression and removal at different stages, ensuring the accuracy of segmentation.

[0135] S500: Identify the interruption points of blood vessels and candidate target points on the vascular skeleton in the initial blood vessel segmentation map.

[0136] Preferably, in order to repair the initial segmentation map The fracture in the middle requires locating the problem point, which includes the following steps:

[0137] 1) Extract the vascular skeleton; for Connectivity is marked to obtain several connected components. For the largest connected region As the main vascular network (based on clinical data, the main retinal vascular network is usually the largest connected region), it is refined to extract its centerline, obtain the skeleton of the main vascular trunk, and construct a topological map G that records information such as nodes (endpoints, intersections), direction, and diameter.

[0138] In this embodiment, the thinning process uses the Zhang-Suen algorithm, and the iteration termination condition is that there is no pixel change in two consecutive iterations.

[0139] 2) Identify the breakpoint P; for non-main trunk lines Perform a neighborhood traversal on other small blood vessel segments; if pixels exist... of Only one blood vessel pixel in the neighborhood is marked as an endpoint; the distance from the endpoint to the center of the optic disc is calculated, and the endpoint closest to the center of the optic disc is selected as the interruption point P of the blood vessel segment.

[0140] Specifically, in this embodiment, the viewing disc is located using grayscale peak values ​​and circularity. Grayscale peak regions are found using a grayscale histogram, and circularity is used for filtering (e.g., circularity). ), determine the viewing area, and take the geometric center of the area as the viewing center.

[0141] 3) Filter candidate target points Q; using each interruption point P as the center, filter all nodes from the vascular skeleton G within its neighboring search range (e.g., 50x50 pixels). And sort them according to their distance from P, selecting the top n as candidate destination points. In this embodiment, n is preferably 5; if there are fewer than 5, all of them are retained.

[0142] Preferably, if there are no candidate target points, the interrupted blood vessel is marked as non-reconstructable and indicated in the final report.

[0143] S600. Perform multi-dimensional verification between the interruption point and each of the candidate target points to determine the target point corresponding to the interruption point and connect them to generate a reconstructed retinal vessel segmentation map.

[0144] In a preferred embodiment, refer to, for example Figure 3 As shown, the multi-dimensional verification, for each interruption point and its corresponding candidate destination point set Perform a three-level progressive verification:

[0145] Level 1 verification: Topological verification based on blood vessel orientation, diameter, or branch angle;

[0146] Secondary verification: Texture feature verification based on pixel grayscale, edge, or gradient features along the connection path;

[0147] Level 3 verification: Specific rules are adapted and verified for special nodes such as blood vessel intersections, branching points, and terminal capillaries.

[0148] More preferably, in a specific embodiment, the topology verification determines whether the blood vessel segment where point P is located is related to... The homology of the vascular segments where the points are located in terms of macroscopic structure is as follows:

[0149] Consistency of direction: Calculate the direction vector of the blood vessel segment where the interruption point is located. The direction vector of the blood vessel segment where the candidate target point is located The angle between To determine whether it is within a reasonable range;

[0150] ;

[0151] In this embodiment, the preferred threshold for achieving consistency is: All angles are vector angles, and the calculation range is... ;

[0152] Diameter matching degree: Calculate the blood vessel diameter at the interruption point. The diameter of the blood vessel at the candidate target point ratio Determine whether it is within a reasonable range:

[0153] ;

[0154] In this embodiment, the preferred threshold for diameter matching degree is... ;

[0155] Branch angle rationality: if the candidate target point As a branch point, the course of the blood vessel is interrupted. Along the main blood vessel The branch angle between them determines the branch angle formed after the connection. Does it conform to the physiological range?

[0156] ;

[0157] In this embodiment, the preferred threshold for the rationality of the branch angle is... .

[0158] More preferably, only when all topology verification items pass, i.e., only when , , When all meet the threshold, candidate points Only then can it proceed to the next level of verification. The above thresholds are all determined based on clinical sample statistics and are consistent with the physiological structural characteristics of retinal vessels.

[0159] More preferably, in one specific embodiment, the texture feature verification verifies from P to... Do the pixels on the potential connection path conform to the microscopic texture features of blood vessels, specifically:

[0160] A texture feature template is established based on the blood vessel segment where point P is located, including grayscale distribution, edge smoothness, gradient direction consistency, etc.

[0161] Search P to The initial path is determined, with the termination condition being reaching point Q. The path can traverse non-vascular regions. The matching degree between pixels along this path and the texture template is calculated. Only when the matching degree is higher than a preset threshold is a candidate point considered. Only then did it pass this level of verification;

[0162] The matching degree includes the degree of fit between the gray value distribution of pixels on the candidate path and the gray value distribution model of the preset blood vessel template; the similarity between the edge smoothness of the candidate path region and the edge smoothness of the preset blood vessel template; and the consistency between the gradient direction of pixels on the candidate path and the preset blood vessel direction vector.

[0163] Specifically, in this embodiment, the texture feature template includes:

[0164] Gray-scale distribution curve: For the K pixels in the blood vessel segment where P is located, the gray-scale values ​​are statistically analyzed. Fitting a normal distribution curve ( The mean, (Standard deviation)

[0165] Edge smoothness: Calculates the standard deviation of the gradient at the edge of the vessel segment. ;

[0166] Gradient direction consistency: statistically correlates pixel gradient directions within a blood vessel segment with the blood vessel orientation vector. The included angle pixel ratio .

[0167] Specifically, the matching degree between pixels and texture templates along this path is calculated. Specifically:

[0168] Search for an initial path L from P to Q (the search terminates when point Q is reached, and the path can pass through non-vascular areas), containing the number of pixels M;

[0169] Gray-scale matching: Calculate the gray-scale value of each pixel in L. and goodness of fit ,statistics number of pixels ;

[0170] Edge matching: Calculate the standard deviation of the edge gradient in region L. ,like If so, it is marked as an edge match;

[0171] Gradient direction matching: Statistical analysis of pixel gradient directions in L and their correlation with the gradient direction of pixels in L. included angle pixel ratio ,like If so, it is marked as gradient direction matching;

[0172] Determine the matching degree for:

[0173] ;

[0174] Among them, when If the verification is successful (the threshold is determined through experimental optimization to balance matching accuracy and computational efficiency), Q is discarded; otherwise, it is discarded.

[0175] Further preferably, in a specific embodiment, specific rules are adapted and verified for special nodes such as vascular intersections, branching points, and terminal capillaries; that is, specific connection judgment rules and / or verification thresholds are used for different types of special nodes. The judgment of all special nodes is based on 5×5 pixel neighborhood analysis.

[0176] Specifically, in this embodiment, the special node includes:

[0177] There exists in the neighborhood For blood vessel segments with different orientations, the angle calculation and topological verification are consistent, both based on the blood vessel orientation vector, and the angles of each segment are... Marked as intersection;

[0178] There is one main blood vessel (diameter) in the neighborhood. ) and 1 to 2 capillaries (diameter) ), marked as branch points;

[0179] Blood vessel diameter at interruption point P Pixels, labeled as capillaries.

[0180] Specifically, in this embodiment, the dedicated rule adaptation includes:

[0181] Intersection rule: Only nodes in the same radial direction as P are selected as Q, and connections across blood vessels with different directions are prohibited;

[0182] Branch point rule: Only "capillaries are allowed to connect to the main blood vessel", that is... For "fine → main" connections, reverse or sibling connections are prohibited;

[0183] Peripheral capillary rule: Appropriately relax the texture verification threshold and narrow the search range. For example, reduce the texture verification threshold to 0.7 and narrow the search range to 30×30 pixels.

[0184] In a preferred embodiment, determining and connecting the destination point corresponding to the interruption point includes:

[0185] For each candidate connection path connecting the interruption point to each candidate destination point, a comprehensive weighted score is calculated. The comprehensive weighted score is obtained by weighting and summing the path length of the candidate connection path and the matching degree obtained in the topology verification and the texture feature verification.

[0186] The candidate connection path with the best comprehensive weighted score is determined as the effective vascular connection path, and the endpoint of the effective vascular connection path is determined as the destination point corresponding to the interruption point.

[0187] A preset curve fitting algorithm is used to smoothly fit and connect along the effective blood vessel connection path.

[0188] More preferably, in a specific embodiment, determining and connecting the destination point corresponding to the interruption point specifically includes:

[0189] 1) For each candidate connection path connecting the interruption point to each candidate destination point, determine the path score, including:

[0190] ;

[0191] ;

[0192] in, Score the path length. For path length, To determine the maximum possible path length within the search range, this embodiment preferably uses 50. To minimize the possible path length, this embodiment preferably uses 10; For topological matching, based on the angle of homologous vessel orientation according to clinical statistics, this embodiment preferably selects the optimal orientation angle as follows: .

[0193] 2) Calculate the comprehensive weighted score of each path, and determine the candidate connection path with the best comprehensive weighted score as the effective vascular connection path:

[0194] ;

[0195] ;

[0196] in, The overall weighted score for candidate connection paths. , , The corresponding terms are weighted; through orthogonal experiments for optimization, this embodiment yields the preferred result. =0.3、 , ;Pick The candidate connection path with the largest value is determined as the effective vascular connection path.

[0197] After three levels of verification, there may still be multiple candidate points Q that are qualified. The most suitable candidate point is determined as the target point by comprehensive weighted score.

[0198] 3) A preset curve fitting algorithm is used to smoothly fit and connect along the effective blood vessel connection path; in this embodiment, a second-order Bezier curve is used to ensure morphological continuity.

[0199] ;

[0200] Wherein, control point M is the intersection of the extended tangent line of the blood vessel segment where interruption point P is located and the extended tangent line of the blood vessel segment where target point Q is located; if the two tangent lines do not intersect, the midpoint of the line connecting P and Q is taken and offset by 5 pixels along the weighted average direction of the two blood vessel direction vectors (the weight is proportional to the blood vessel diameter), ensuring The difference between the curvature and the original vascular curvature .

[0201] Preferably, after the connection is completed, verification is performed, specifically as follows:

[0202] The connecting segment has normal connectivity with the original blood vessel and is free from tortuosity (changes in curvature). If any breakpoint is found, retain it; otherwise, search again. After connecting all breakpoints, output the final, complete, and coherent reconstructed retinal vessel segmentation map. .

[0203] In a preferred embodiment, the following steps are also included:

[0204] Based on the reconstructed retinal vessel segmentation map, the vessel segments are distinguished as arteries or veins, and arterial and vein markers are generated;

[0205] Quantify at least one vascular morphology parameter, including vascular tortuosity, branching angle, or arteriovenous diameter ratio, wherein the arteriovenous diameter ratio is calculated based on the ratio of the average diameter of the artery to the average diameter of the vein measured in a preset area;

[0206] Generate quantitative analysis results that include arterial and venous markers and the morphological parameters.

[0207] Specifically, in this embodiment, arterial / venous (A / V) distinction is performed on the vascular segments. Specifically, the reconstructed vascular tree is segmented into independent vascular segments, and features such as the mean grayscale value, width, and branch density of each segment are extracted. These features are then input into a pre-trained support vector machine (SVM) classifier to automatically label arteries and veins. The branch density is the number of branch points per unit area (unit: ...). ).

[0208] The pre-trained Support Vector Machine (SVM) classifier described in this embodiment is trained using 1000 labeled fundus images of arteries and veins, with an artery to vein sample ratio of 1:1 and a training set to test set ratio of 7:3. The classification features are determined by mutual information entropy filtering to ensure optimal correlation between the features and the differences between arteries and veins.

[0209] More preferably, the quantitative analysis results also include abnormal indicator prompts (such as AVR < 0.7).

[0210] The retinal vessel segmentation device in the fundus image according to the second embodiment of this application is used to perform the method described in the first embodiment, as referred to below. Figure 4 As shown, it includes:

[0211] The analysis module is configured to analyze the acquired fundus images to determine at least one prior analysis index characterizing the distribution characteristics of the images, and to identify noise regions in the images to generate noise classification information.

[0212] The image processing module is configured to dynamically determine processing parameters based on the prior analysis indicators and the noise classification information; it is also configured to process the fundus image using the processing parameters and perform noise removal processing based on the noise classification information to generate an initial vascular segmentation map containing interrupted blood vessels.

[0213] The intelligent reconstruction module is configured to identify the interruption points of blood vessels and candidate target points on the vascular skeleton in the initial blood vessel segmentation map; and to perform multi-dimensional verification between the interruption points and each candidate target point to determine the target point corresponding to the interruption point and connect them to generate the reconstructed retinal blood vessel segmentation map.

[0214] It should be noted that the retinal vessel segmentation device in fundus images provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of this application can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of this application are only for distinguishing the various modules or steps and are not considered as an improper limitation of this application.

[0215] A device according to a third embodiment of this application includes:

[0216] At least one processor;

[0217] and a memory communicatively connected to at least one of the processors;

[0218] The memory stores instructions that can be executed by the processor to implement the above-described method for segmenting retinal vessels in fundus images.

[0219] A computer-readable storage medium according to a fourth embodiment of this application stores computer instructions for execution by a computer to implement the above-described method for segmenting retinal vessels in fundus images.

[0220] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0221] The following is for reference. Figure 5 It shows a schematic diagram of the structure of a computer system for implementing embodiments of the systems, methods, and electronic devices of this application. Figure 5 The server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0222] like Figure 5 As shown, the computer system includes a Central Processing Unit (CPU) 501, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 502 or programs loaded from storage section 508 into Random Access Memory (RAM) 503. The RAM 503 also stores various programs and data required for system operation. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An Input / Output (I / O) interface 505 is also connected to the bus 504.

[0223] The following components are connected to I / O interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to I / O interface 505 as needed. Removable media 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 510 as needed so that computer programs read from them can be installed into storage section 508 as needed.

[0224] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by central processing unit (CPU) 501, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0225] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0226] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0227] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0228] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0229] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for segmenting retinal vessels in fundus images, characterized in that, Includes the following steps: Acquire the fundus image to be processed; The fundus image is analyzed to determine at least one prior analysis index characterizing the distribution characteristics of the image, and noise regions in the image are identified to generate noise classification information; Based on the prior analysis indicators and noise classification information, the processing parameters used to extract blood vessel segmentation features are dynamically determined; The fundus image is processed using the dynamically determined processing parameters, and noise removal is performed based on the noise classification information to generate an initial vascular segmentation map containing interrupted blood vessels. Identify the interruption points of blood vessels and candidate target points on the vascular skeleton in the initial blood vessel segmentation image; Multi-dimensional verification is performed between the interruption point and each of the candidate target points to determine the target point corresponding to the interruption point and connect them to generate a reconstructed retinal vessel segmentation map.

2. The method for segmenting retinal vessels in fundus images according to claim 1, characterized in that, It also includes preprocessing of the fundus images to be processed, specifically including: Obtain the green channel image of the fundus; The green channel image is filtered using a mean filtering algorithm; An adaptive histogram equalization algorithm was used to enhance the contrast of the filtered green channel image to obtain a preprocessed fundus image.

3. The method for segmenting retinal vessels in fundus images according to claim 1, characterized in that, The prior analysis indicators include at least one of the following: Blood vessel density obtained from fundus images based on sliding window statistics; The noise intensity is calculated by the ratio of the global grayscale variance to the mean edge gradient of the fundus image; Gray-level uniformity is calculated by the skewness of the gray-level distribution in fundus images.

4. The method for segmenting retinal vessels in fundus images according to claim 3, characterized in that, The processing parameters include at least the window size for calculating the Hesser matrix and a feature selection threshold for screening vascular features; the dynamic determination of the processing parameters for extracting vascular segmentation specifically includes: The initial window size is matched based on the blood vessel density and the noise intensity; An initial feature screening threshold is matched based on the noise intensity and the gray-level uniformity. Based on the noise classification information, the initial window size and initial feature selection threshold for different types of noise regions are adaptively adjusted.

5. The method for segmenting retinal vessels in fundus images according to claim 4, characterized in that, The generation of the initial vascular segmentation map containing interrupted blood vessels includes: Based on the dynamically determined processing parameters, adaptive Heather matrix feature extraction is performed on the fundus image to generate a vascular feature map; The vascular feature map is binarized to generate a binary vascular map; Based on the noise classification information, type-specific removal is performed on pseudo-vascular targets corresponding to the noise region in the binary vascular map to generate the initial vascular segmentation map.

6. The method for segmenting retinal vessels in fundus images according to claim 5, characterized in that, Before binarizing the vascular feature map, feedback calibration is also included, specifically including: Calculate the vascular continuity in the vascular feature map, which represents the proportion of vascular pixels within an effective connected region, and the noise false detection rate, which represents the proportion of isolated suspected vascular pixels. When the continuity of the blood vessels is lower than a preset lower limit, the feature selection threshold is lowered and adaptive Hesser matrix feature extraction is re-executed; when the false detection rate of noise is higher than a preset upper limit, the feature selection threshold is increased and adaptive Hesser matrix feature extraction is re-executed; until the preset quality requirements are met.

7. The method for segmenting retinal vessels in fundus images according to claim 1, characterized in that, The multi-dimensional verification includes: Topological verification based on blood vessel orientation, diameter, or branch angle; Texture feature verification based on pixel grayscale, edge, or gradient features along the connection path; Specific rules are adapted and validated for special nodes such as blood vessel intersections, branching points, and terminal capillaries.

8. The method for segmenting retinal vessels in fundus images according to claim 7, characterized in that, The step of determining the destination point corresponding to the interruption point and connecting it includes: For each candidate connection path connecting the interruption point to each candidate destination point, a comprehensive weighted score is calculated. The comprehensive weighted score is obtained by weighting and summing the path length of the candidate connection path and the matching degree obtained in the topology verification and the texture feature verification. The candidate connection path with the best comprehensive weighted score is determined as the effective vascular connection path, and the endpoint of the effective vascular connection path is determined as the destination point corresponding to the interruption point. A preset curve fitting algorithm is used to smoothly fit and connect along the effective blood vessel connection path.

9. The method for segmenting retinal vessels in fundus images according to claim 1, characterized in that, The method further includes: Based on the reconstructed retinal vessel segmentation map, the vessel segments are distinguished as arteries or veins, and arterial and vein markers are generated; Quantify at least one vascular morphology parameter, including vascular tortuosity, branching angle, or arteriovenous diameter ratio, wherein the arteriovenous diameter ratio is calculated based on the ratio of the average diameter of the artery to the average diameter of the vein measured in a preset area; Generate quantitative analysis results that include arterial and venous markers and the morphological parameters.

10. A device for segmenting retinal vessels in fundus images, characterized in that, include: The analysis module is configured to analyze the acquired fundus images to determine at least one prior analysis index characterizing the distribution characteristics of the images, and to identify noise regions in the images to generate noise classification information. The image processing module is configured to dynamically determine processing parameters based on the prior analysis indicators and the noise classification information; it is also configured to process the fundus image using the processing parameters and perform noise removal processing based on the noise classification information to generate an initial vascular segmentation map containing interrupted blood vessels. The intelligent reconstruction module is configured to identify the interruption points of blood vessels and candidate target points on the vascular skeleton in the initial blood vessel segmentation map; and to perform multi-dimensional verification between the interruption points and each candidate target point to determine the target point corresponding to the interruption point and connect them to generate the reconstructed retinal blood vessel segmentation map.

Images