A feature image extraction method for retinal vascular morphological analysis

Abstract

This invention relates to the field of feature image extraction, and more particularly to a feature image extraction method for retinal vessel morphology analysis. The method includes: acquiring a raw retinal color image and preprocessing it to obtain a preprocessed grayscale image; constructing a double-boundary annular neighborhood based on the preprocessed grayscale image; uniformly dividing the double-boundary annular neighborhood, calculating sector integrals, and selecting the optimal branch direction pairs; calculating the retinal vessel branch structure response values ​​by coupling the intensity balance and local-global intensity of the optimal branch direction pairs, and selecting candidate regions for vessel branches to obtain an initial binary image; performing morphological operations, skeleton extraction, and spur removal on the initial binary image to obtain a retinal vessel morphology feature map. This method solves the problems of low branch recognition accuracy, weak anti-interference ability, discontinuous vessel morphology, susceptibility to background and noise interference, and difficulty in accurately distinguishing single vessels from branch regions.

Description

Technical Field

[0001] This invention relates to the field of feature image extraction, and more particularly to a feature image extraction method for retinal vascular morphology analysis. Background Technology

[0002] The morphology, distribution, and branching structure of retinal vessels are core imaging evidence for ophthalmic disease screening, disease progression monitoring, and prognostic assessment. Accurate extraction of vascular branch points and continuous vascular skeletons plays a crucial supporting role in clinical pathological analysis and diagnostic decision-making. Current retinal vessel extraction technology has formed a relatively complete system. Mainstream techniques primarily utilize global grayscale enhancement, fixed threshold segmentation, and traditional edge detection operators to achieve basic vessel contour extraction. Some improved schemes further enhance linear vessel features and improve the recognition of main vessels by introducing multi-scale analysis and directional filtering. Cutting-edge research relies on deep learning network models, using massive amounts of data to optimize segmentation results, attempting to overcome the shortcomings of traditional algorithms in terms of completeness. Various techniques demonstrate stable performance in the segmentation and recognition of straight single-branch vessel segments, and can initially outline the contours of the main retinal vessels, making them suitable for basic analysis scenarios in routine fundus images and possessing certain practical applications in general vessel detection.

[0003] However, the above-mentioned technologies generally have the following specific defects: for branching areas such as bifurcations and intersections of blood vessels, the positioning accuracy and anti-interference ability are seriously insufficient, and problems such as false detection, missed detection, blood vessel rupture, and spiculation artifacts are prone to occur; although deep learning models improve the overall integrity, they have shortcomings such as high computational complexity, sensitivity to uneven lighting and noise, weak differentiation of branching areas, and inability to simultaneously and accurately extract small microvessels and main blood vessels, making it difficult to meet the core clinical needs for rapid, stable and accurate detection of blood vessel branches. Summary of the Invention

[0004] This invention provides a feature image extraction method for retinal vascular morphology analysis, which solves the technical problems of traditional retinal feature image extraction methods, such as low branch recognition accuracy, weak anti-interference ability, discontinuous vascular morphology, susceptibility to background and noise interference, and difficulty in accurately distinguishing single vessels from branch regions.

[0005] The present invention provides a method for feature image extraction for retinal vascular morphology analysis, comprising the following steps: S1. Acquire the original retinal color image, and sequentially perform grayscale processing, background correction processing and preliminary blood vessel enhancement processing to obtain the preprocessed grayscale image; S2. Based on the preprocessed grayscale image, a double-boundary annular neighborhood is constructed; after uniformly dividing the double-boundary annular neighborhood, the sector integral is calculated, and the optimal branch direction pair is selected; by coupling the intensity balance and local global intensity of the optimal branch direction pair, the response value of the retinal vascular branch structure is calculated, and the candidate regions of vascular branches are selected to obtain the initial binary map; morphological operations, skeleton extraction and spur removal are performed on the initial binary map to obtain the retinal vascular morphological feature map.

[0006] Preferably, S1 specifically includes: The original retinal color image is converted to grayscale to obtain a grayscale image; the grayscale image is then subjected to background correction through low-pass filtering iteration to obtain a background-corrected grayscale image; and the background-corrected grayscale image is then subjected to preliminary vascular enhancement through multi-directional, multi-scale Gaussian difference filtering to obtain a preprocessed grayscale image.

[0007] Preferably, S2 specifically includes: The double-boundary annular neighborhood is uniformly divided, and the central angle of the sector is calculated. Based on the central angle of the sector, the minimum included angle of any candidate sector pair is calculated. Candidate sector pairs that are not adjacent in the circumferential direction and whose included angles are not collinear or opposite are selected, and the combination with the largest sum of sector integrals is selected as the optimal branch direction pair.

[0008] Preferably, S2 specifically includes: Based on the optimal branch direction pair, a directional balance term and a global intensity term are introduced to construct the response formula of the retinal vascular branch structure and calculate the response value of the retinal vascular branch structure.

[0009] Preferably, S2 specifically includes: By calculating the product of the sector integrals in the two directions of the optimal branch direction pair and the sum of the squares of the sector integrals, a directional equilibrium term is constructed to characterize the degree of response equilibrium of the retinal vascular branch structure in the two optimal branch directions.

[0010] Preferably, S2 specifically includes: Based on the sector integral sum of the two directions of the optimal branch direction pair, combined with the logarithmic function, a global intensity term is constructed to characterize the local overall response intensity.

[0011] Preferably, S2 specifically includes: The response values ​​of retinal vascular branch structures are calculated pixel by pixel, and a branch structure response map is generated. The branch structure response map is then thresholded to determine the candidate regions of vascular branches, and an initial binary map is obtained.

[0012] Preferably, S2 specifically includes: Morphological opening operations are performed on the initial binary image to obtain the effective vascular structure region. A morphological iterative thinning algorithm with 8-neighborhood connectivity rules and endpoint protection is used to extract the effective vascular structure region into a continuous slender skeleton with a single pixel width. After spur removal, the retinal vascular morphological feature map is obtained.

[0013] The beneficial effects of the technical solution of the present invention are: 1. A three-level preprocessing workflow of grayscale conversion, background correction, and vessel enhancement is adopted to eliminate interference from color, lighting, and background fluctuations step by step, improve the contrast between blood vessels and the background, and provide a high-quality input foundation for the subsequent accurate detection of branch structures.

[0014] 2. A ring sector integration and direction filtering mechanism is designed, which combines non-collinear sector to make dual judgments on balance and local response intensity. It can accurately distinguish between vascular branch points, ordinary vascular segments and noisy areas. It has high branch positioning accuracy, strong anti-interference ability and is suitable for retinal images of different resolutions.

[0015] 3. Through a coherent post-processing workflow of threshold screening, connected component denoising, morphological refinement, and spur removal, while preserving the vascular topology and branching direction, spurious responses and minor interferences are removed, ultimately resulting in a continuous, smooth, and accurate retinal vascular morphological feature map, which can directly support ophthalmic clinical diagnosis, vascular parameter quantification, and disease follow-up analysis. Attached Figure Description

[0016] Figure 1 This is a flowchart of a feature image extraction method for retinal vascular morphology analysis according to the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following describes in detail, with reference to the accompanying drawings, a specific scheme for a feature image extraction method for retinal vascular morphology analysis provided by the present invention.

[0020] See attached document Figure 1The diagram illustrates a flowchart of a feature image extraction method for retinal vascular morphology analysis according to an embodiment of the present invention. The method includes the following steps: S1. Acquire the original retinal color image, and sequentially perform grayscale processing, background correction processing and preliminary blood vessel enhancement processing to obtain the preprocessed grayscale image; First, raw retinal color images are obtained using fundus examination data from clinical medical institutions and publicly available retinal image datasets. These raw retinal color images contain information such as retinal vessels, optic disc, background, and illumination noise. To eliminate the impact of color channel differences on vessel feature extraction, the raw retinal color images are converted to grayscale using a weighted average method, transforming the three-channel color retinal image into a single-channel grayscale image to reduce interference from differences in different color channels during subsequent vessel feature extraction. After obtaining the grayscale image, background correction processing is performed. Specifically, a low-pass filtering iterative algorithm is used to correct the grayscale image. The image undergoes multiple iterations of global smoothing to filter out high-frequency signals such as blood vessels and noise, retaining only the background grayscale information of avascular and noise-free areas of the retina. This yields the background grayscale distribution of the entire grayscale image. Then, the grayscale value of each pixel in the grayscale image is subtracted pixel-by-pixel from the corresponding background grayscale value to homogenize the overall image brightness, significantly reducing interference from background undulations, resulting in a background-corrected grayscale image. After background correction, to highlight the slender, strip-like dark streaks of retinal blood vessels, multi-directional, multi-scale Gaussian difference filtering is further applied to the background-corrected grayscale image for initial vascularization. The enhancement process involves: constructing a two-dimensional Gaussian difference kernel, where the small-scale standard deviation is 1.0~1.5 pixels (adapting to tiny blood vessels with a diameter of 1-2 pixels), and the large-scale standard deviation is 2.0~2.5 pixels (adapting to blood vessel branches with a diameter of 2-3 pixels). This generates a linear filter kernel with a major axis of 11 pixels and a minor axis of 3 pixels, which is then rotated to four directions: 0°, 45°, 90°, and 135°. Using these four linear filter kernels, the background-corrected grayscale image is filtered to obtain filter response maps for each direction and scale. The maximum response value is then used for each pixel location. The image is fused by inverting the phase to enhance the vascular dark texture features and setting a low threshold to truncate and remove weak noise responses, ultimately resulting in a preprocessed grayscale image with clear vascular dark textures, sharp boundaries, a clean background, and suppressed noise. The low threshold is obtained by statistically calibrating multiple retinal images of the same resolution at multiple scales. Specifically, multiple retinal images of the same resolution as the image to be processed are selected, and after background correction and multi-directional Gaussian difference filtering, the grayscale histogram of the filtered response map is statistically analyzed to determine the noise response concentration interval and the vascular response start interval, and the median between the upper limit of the noise response and the lower limit of the vascular response is taken as the low threshold.

[0021] S2. Based on the preprocessed grayscale image, a double-boundary annular neighborhood is constructed; after uniformly dividing the double-boundary annular neighborhood, the sector integral is calculated, and the optimal branch direction pair is selected; by coupling the intensity balance and local global intensity of the optimal branch direction pair, the response value of the retinal vascular branch structure is calculated, and the candidate regions of vascular branches are selected to obtain the initial binary map; morphological operations, skeleton extraction and spur removal are performed on the initial binary map to obtain the retinal vascular morphological feature map.

[0022] After obtaining the preprocessed grayscale image, a double-boundary annular neighborhood is constructed with the pixel to be detected as the center. The annular neighborhood consists of an inner radius... and outer radius Determined jointly, and satisfying Based on the physiological size characteristics of retinal vessel branches (the diameter of tiny retinal microvessels is approximately 1-3 pixels, and the local area of ​​the branch region usually does not exceed 10 pixels), and considering the requirement that the annular neighborhood needs to cover the local area of ​​the branch, the inner radius is determined. Take 3 to 5 pixels, outer radius Select pixels 7 to 11. After establishing the annular neighborhood, divide the annular neighborhood evenly along the central angle. A fixed sector, the The preset even number is determined based on the theory of balancing directional resolution and computational efficiency, combined with the physiological characteristic that retinal vessels mostly branch in 2-4 directions. As a specific embodiment, a reference value of 16 is used. Since the annular neighborhood is uniformly divided, the angular width of a single sector is... Let the first The central angle of each sector is The formula for calculating the central angle Confirmed, among which , is the sector index.

[0023] Subsequently, the sector integral is calculated for each sector using a grayscale value accumulation method. The magnitude of the sector integral directly reflects the strength of the vascular structure in the corresponding sector direction. After obtaining the integrals of all sectors, for any pair of candidate sectors, the difference in central angle between the candidate sector pairs is first calculated, and then... Periodic normalization is mapped to an interval To obtain the minimum included angle Furthermore, candidate sector pairs that simultaneously meet the following conditions are selected and retained as candidate branch direction pairs: first, the two sectors do not share a common boundary in the circumferential direction, so as to exclude local continuous edge responses caused by adjacent directions; second, the minimum included angle... satisfy To further exclude collinear opposite directions and non-branching wide edge responses; among all candidate branch direction pairs that meet the screening criteria, the sum of the two sector integrals is sorted from largest to smallest, and the sector pair with the largest sum of sector integrals is selected as the optimal branch direction pair.

[0024] After obtaining the optimal branch direction pair, a response formula for the retinal vascular branch structure is constructed based on the local directional response equilibrium theory in digital image processing and the geometric topological characteristics of retinal vascular branches. This response formula, by introducing a coupling mechanism between a directional equilibrium term and a global intensity term, balances the equilibrium of bidirectional vascular responses while quantifying the overall local vascular intensity, thus achieving precise differentiation between branched and non-branched regions. The retinal vascular branch structure response formula is as follows: ; in, Indicated in pixels Response values ​​of retinal vascular branch structures at the location; , These represent the horizontal and vertical coordinates in the preprocessed grayscale image, respectively. To prevent extremely small positive numbers with a denominator of zero, the value can be taken as... ; Indicated in pixels Within the annular neighborhood of the point, the sector integral value with the largest centering value of the optimal branch direction is used to reflect the vascular strength in the local first principal extension direction; Indicated in pixels Within the annular neighborhood of the point, the smaller sector integral value of the optimal branch direction is used to reflect the vascular strength in the local second principal extension direction; The directional balance term is used to characterize the degree of response balance of retinal vascular branch structures in two optimal branch directions. When the sector integrals in both directions are strong and relatively close, the value of the directional balance term is large. When there is only a single main direction extension, or when the response difference of retinal vascular branch structures in two optimal branch directions is large, the value of the directional balance term is reduced, which can suppress the false response of ordinary single-branch segments, wide vessel edges and pseudo-branch regions. The global intensity term is used to characterize the intensity of the local overall response. When there are indeed two relatively obvious non-collinear blood vessels extending within the annular neighborhood, the global intensity term increases with the enhancement of the overall response.

[0025] Calculate the response value of the retinal vascular branch structure for each pixel in the preprocessed grayscale image. The branch structure response map is composed of the response values ​​of the retinal vascular branch structures at all pixel locations. Then analyze the response diagram of the branch structure. Threshold filtering is performed to identify connected regions composed of pixels whose retinal vascular branch structure response values ​​are greater than or equal to the response threshold as candidate regions for vascular branches, and to identify and remove connected regions composed of pixels whose retinal vascular branch structure response values ​​are less than the response threshold as background regions, thus completing the initial separation of vascular branches from the background and obtaining an initial binary image. The response threshold is obtained through a calibration stage. Specifically, for multiple retinal images at the same image resolution level, the corresponding branch structure response map is first calculated. Then, the Otsu method is used to calculate the global threshold for each branch structure response map to obtain a candidate threshold set. The candidate threshold set is then statistically processed, and the mean value is taken as the response threshold. Next, morphological operations are performed on the initial binary image. Through opening operations (erosion followed by dilation), isolated small-area noise regions and unnecessary short pseudo-branches in the initial binary image can be removed, while retaining a large range of vascular structures, resulting in an effective vascular structure region. Then, the Zhang-Suen morphological iterative refinement algorithm is used to extract the skeleton of the effective vascular structure region. The specific implementation process is as follows: First, the effective vascular structure region obtained after morphological opening operations is used as input. The vascular region is set as a white foreground (grayscale value 255), and the background region is set as... The connectivity of the blood vessel region is defined using an 8-neighborhood connectivity rule, with black as the base color (grayscale value 0). Then, non-critical pixels at the vessel edges are alternately deleted through two rounds of iterative sub-processes in a fixed order. The first iteration specifically deletes boundary points meeting the removal criteria in the east, west, south, and north directions. The second iteration deletes boundary points meeting the removal criteria in the southeast, northeast, southwest, and northwest directions. The removal criteria are: during the iteration process, only endpoint pixels with only one foreground pixel in their neighborhood are retained to avoid breakage at the vessel tip. This iterative process is repeated until no pixels meeting the removal criteria can be deleted. Finally, while preserving the topological connectivity of the blood vessel and maintaining the branch direction throughout, the wide blood vessel region is refined into a continuous, slender skeleton with a single pixel width. To address the interference from short spurs and pseudo-branches that easily appear in the refined skeleton, fine spur structures with pixel lengths below a preset length threshold are further removed, and redundant interference is cleaned up, ultimately resulting in a continuous, smooth, and spur-free retinal vascular morphological feature map. The length threshold is obtained through calibration using retinal image samples of the same resolution. Specifically, multiple retinal images of the same resolution as the image to be processed are selected. After preprocessing, morphological opening operations, and skeleton extraction, the maximum pixel length of pseudo-spurs and short pseudo-branches is manually measured against the minimum pixel length of the actual small blood vessel terminals. The threshold range is determined to be between the two. The statistical average of the threshold ranges of multiple sets of retinal images is then taken to obtain the length threshold.

[0026] In summary, a feature image extraction method for retinal vascular morphology analysis has been developed.

[0027] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0028] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0029] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A feature image extraction method for retinal vascular morphological analysis, characterized by, The method comprises the following steps: S1. Obtain an original retinal color image, and sequentially perform grayscale processing, background correction processing and preliminary blood vessel enhancement processing to obtain a preprocessed grayscale image; S2. Based on the preprocessed grayscale image, a double-boundary annular neighborhood is constructed; After uniformly dividing the double-boundary annular neighborhood, sector integrals are calculated, and the optimal branch direction pair is selected; the retinal blood vessel branch structure response value is calculated by coupling the intensity balance and the local-global intensity of the optimal branch direction pair, and the blood vessel branch candidate region is selected to obtain an initial binary image; morphological operation, skeleton extraction and burr removal are performed on the initial binary image to obtain a retinal blood vessel morphological feature map.

2. The feature image extraction method for retinal vascular morphological analysis according to claim 1, wherein, The S1 specifically comprises: Grayscale processing is performed on the original retinal color image to obtain a grayscale image; background correction processing is performed on the grayscale image by low-pass filtering iteration to obtain a background-corrected grayscale image; preliminary blood vessel enhancement processing is performed on the background-corrected grayscale image by multi-direction and multi-scale Gaussian difference filtering to obtain a preprocessed grayscale image.

3. The feature image extraction method for retinal vascular morphological analysis according to claim 1, wherein, The S2 specifically comprises: The double-boundary annular neighborhood is uniformly divided, and the central angles of the sectors are calculated; based on the central angles of the sectors, the minimum included angle of any candidate sector pair is calculated; the candidate sector pairs that are not adjacent in the circumferential direction and have non-collinear and non-reversed included angles are selected, and the combination with the maximum sector integral and is selected as the optimal branch direction pair.

4. The feature image extraction method for retinal vascular morphological analysis according to claim 3, wherein, The S2 specifically comprises: Based on the optimal branch direction pair, a direction balance term and a global intensity term are introduced to construct a retinal blood vessel branch structure response formula, and the retinal blood vessel branch structure response value is calculated.

5. The feature image extraction method for retinal vascular morphological analysis according to claim 4, wherein, The S2 specifically comprises: The product of the sector integrals in the two directions of the optimal branch direction pair and the square sum of the sector integrals are calculated to construct the direction balance term, which represents the response balance degree of the retinal blood vessel branch structure of the two optimal branch directions.

6. The feature image extraction method for retinal vascular morphological analysis according to claim 4, wherein, The S2 specifically comprises: Based on the sector integral sum of the two directions of the optimal branch direction pair, a global intensity term is constructed by combining a logarithmic function, which represents the local overall response intensity.

7. The feature image extraction method for retinal vascular morphological analysis according to claim 4, wherein, The S2 specifically comprises: The retinal blood vessel branch structure response value is calculated pixel by pixel, and a branch structure response map is generated; the branch structure response map is threshold filtered to determine the blood vessel branch candidate region, and an initial binary image is obtained.

8. The feature image extraction method for retinal vascular morphological analysis according to claim 7, wherein, The S2 specifically comprises: Morphological opening operation is performed on the initial binary image to obtain an effective blood vessel structure region; an 8-neighborhood connectivity rule and an endpoint protection morphological iterative thinning algorithm are used to extract the effective blood vessel structure region into a single-pixel-width continuous slender skeleton, and burr removal is performed to obtain a retinal blood vessel morphological feature map.

Images