A method and apparatus for extracting retinal feature parameters based on optical coherence tomography (OCT) angiography.
By combining the Frangi vascular filter of the Hessian matrix and the Sauvola thresholding strategy, OCTA images are preprocessed, solving the image quality problem of OCTA. This enables accurate extraction and multi-level analysis of retinal vascular and texture feature parameters, and is applicable to different devices and pathological conditions, supporting early screening and auxiliary diagnosis of fundus diseases.
Patent Information
- Application Number
- CN202611303981.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
Existing OCTA images are susceptible to factors such as speckle noise, insufficient contrast, and projection artifacts during acquisition, resulting in blurred vessel edges and discontinuous small branches, which affects the accuracy and stability of vascular feature parameter calculation. Furthermore, the generalization ability of existing methods is limited under different equipment and pathological conditions.
A method combining the Frangi vessel filter based on the Hessian matrix with the Sauvola thresholding strategy is used to preprocess OCTA images, including normalization, vessel enhancement, local threshold binarization, and region segmentation, and to calculate retinal vessel and texture feature parameters.
It improves the accuracy and stability of vascular and texture feature parameter extraction, realizes the joint extraction of multiple feature parameters, supports the comprehensive analysis of multi-level vascular networks, and has good cross-device applicability and computational efficiency.
Smart Images

Figure CN123118513A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer image processing technology, and in particular to a method and apparatus for extracting retinal vascular feature parameters and texture feature parameters based on optical coherence tomography angiography (OCTA). Background Technology
[0002] The retinal vascular network is a crucial component of the retina, and its morphology, distribution, and topology comprehensively reflect the functional state and pathological changes of the retinal microcirculation system. Because retinal vessels can be directly and non-invasively observed, their network structure not only reflects local blood perfusion in the eye but also, to some extent, reflects the overall health of the body's microvascular system. The course, branching patterns, diameter variations, tortuosity, and density of retinal vessels are all closely related to the physiological function of the retina and provide important evidence for the early screening, auxiliary diagnosis, and disease assessment of various fundus diseases and systemic diseases (such as diabetic retinopathy, glaucoma, and age-related macular degeneration).
[0003] OCTA is a non-invasive, rapid, tomographic optical imaging technique capable of visualizing the microvascular structures in the retina and choroid. The basic principle of OCTA is to identify dynamic scattering signals caused by the movement of red blood cells within vessels by detecting changes in the intensity or phase characteristics of OCT signals in multiple B-scans acquired repeatedly over a specific time period, thereby reconstructing the retinal microvascular network. In contrast, the OCT signals corresponding to static tissues typically remain relatively stable over time. The obtained OCTA signal is related to the motion state of flowing red blood cells; within a certain range, the faster the blood flow velocity, the stronger the flow signal, until saturation is reached.
[0004] OCTA (Optical Characteristic Amplification) can display retinal microvascular structures at high resolution, making it an important image source for quantitative analysis of retinal vessels. Based on OCTA images, various parameters reflecting vascular structure and distribution characteristics can be obtained, including fractal dimension (FD), vessel density (VD), skeleton density (SKD), and vessel diameter index (VDI), providing a basis for an objective description of the retinal microvascular network. However, during actual acquisition, OCTA images are often affected by factors such as speckle noise, insufficient contrast, projection artifacts, motion artifacts, and local signal attenuation, leading to blurred vessel edges, unclear capillary display, discontinuous small branches, and even local vessel loss or pseudo-structural enhancement in the images.
[0005] In the aforementioned situations, directly performing vascular analysis based on raw or low-quality OCTA images can easily lead to inaccurate vascular region identification, thereby affecting the calculation results of vascular feature parameters. For example, blurred vascular boundaries reduce the accuracy of vascular diameter measurement, incomplete display of small vessels affects vascular density and length statistics, and impaired vascular connectivity may affect the accuracy of calculations for parameters such as branch number, skeletal structure, and tortuosity. Therefore, the accuracy of retinal vessel segmentation and the fidelity of microvascular structure are directly related to the accuracy and stability of vascular feature parameter extraction results.
[0006] Existing methods for OCTA vessel segmentation mainly include traditional image processing methods, machine learning methods based on artificial features, and deep learning-based segmentation methods. While deep learning-based vessel segmentation methods have achieved good results in terms of segmentation accuracy, they typically rely on large amounts of high-quality labeled data. In practice, OCTA image vessel annotation is costly, highly subjective, and inconsistent among different annotators. Furthermore, existing models are still prone to vessel breakage, incomplete segmentation, or misidentification of artifacts as real vessels when dealing with small capillaries, low-contrast regions, areas with weak blood flow signals, and projection artifacts. Methods based on artificial features and traditional classifiers are highly dependent on feature extraction methods and empirical design, making it difficult to fully represent the multi-scale, fine structure, and complex topological relationships of vessels in OCTA images. Therefore, their generalization ability is limited under different devices, scanning parameters, or pathological conditions.
[0007] Traditional methods such as thresholding, filtering enhancement, and morphological processing, while often relying on manually designed rules and parameter settings, are sensitive to image grayscale variations, noise interference, and scan quality fluctuations. They struggle to adapt to the complexities commonly found in OCTA images, such as speckle noise, uneven contrast, projection artifacts, and minor vascular breaks, leading to issues like missed detections, false detections, and discontinuous vessel boundaries. However, traditional OCTA vessel segmentation methods offer advantages such as simple algorithm implementation, low computational complexity, minimal dependence on training data, and strong interpretability of results, making them practically valuable in computationally limited applications. Some traditional methods can also be used as preprocessing modules to enhance vascular structure, suppress noise interference, or improve the connectivity of segmentation results.
[0008] It is worth noting that although various vascular enhancement techniques and thresholding strategies have been used in OCTA image processing, existing research lacks a systematic comparison of these methods, and the impact of different method combinations on the extraction results of vascular feature parameters is still not quantitatively evaluated. Studies have shown that the calculated vascular density (VD) is highly dependent on the selected vascular enhancement technique and thresholding strategy, and different method combinations may yield significantly different VD values from the same OCTA image. Therefore, determining the optimal combination of image processing methods is crucial for achieving accurate and stable extraction of retinal vascular feature parameters.
[0009] Furthermore, texture features are mathematical features that quantify the spatial distribution of pixel gray levels in an image and are widely used in medical image analysis. Texture features extract visual patterns and structural features that are difficult for the human eye to directly recognize through computerized texture analysis of images. Typical texture features include contrast, correlation, entropy, energy, and homogeneity. Studies have shown that texture feature analysis based on OCT images can identify early changes in the retina during the "silent" period of disease. Changes in texture features occur concurrently with biological changes in the retina, providing new possibilities for the early diagnosis of diabetic retinopathy (DR).
[0010] However, OCTA images are often susceptible to degradation in quality during actual acquisition due to factors such as speckle noise, insufficient contrast, projection artifacts, motion artifacts, and local signal attenuation. If texture feature extraction is performed directly on raw or low-quality OCTA images, noise and artifacts in the image will interfere with the accurate representation of texture features, affecting the accuracy and stability of texture feature calculation. Therefore, appropriate preprocessing is necessary before extracting texture features from OCTA images to enhance vascular structure and suppress noise interference. Summary of the Invention
[0011] The purpose of this invention is to provide a method and apparatus for extracting retinal vascular feature parameters and texture feature parameters based on OCTA images. By systematically optimizing the combination of vascular enhancement technology and threshold strategy, it achieves accurate segmentation of the multilayer retinal vascular network and reliable extraction of various vascular features and texture feature parameters, providing technical support for early screening and auxiliary diagnosis of fundus diseases.
[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0013] On one hand, the present invention provides a method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging, which includes the following steps:
[0014] Step 1, OCTA Image Acquisition and Normalization Preprocessing: Receive fundus OCTA images and perform normalization preprocessing on the received fundus images to obtain preprocessed images;
[0015] Step 2: Perform vessel enhancement processing on the normalized OCTA image using the Frangi vessel filter based on the Hessian matrix;
[0016] Step 3: Binarization using a thresholding strategy: The Sauvola local thresholding strategy is used to binarize the image after vascular enhancement.
[0017] Step 4: Binarized image region segmentation: Based on the size of the received fundus OCTA image, the region is divided into multiple regions of interest;
[0018] Step 5: Calculation of retinal feature parameters: After obtaining the binarized fundus image, calculate the retinal vascular feature parameters and texture feature parameters for each region;
[0019] Step 6: Repeat steps 1 to 5 to extract vascular feature parameters and texture feature parameters from the six layered images of the retina.
[0020] Furthermore, in step 1, the original image grayscale values are mapped to the [0,1] range through normalization processing to eliminate the impact of differences in image grayscale range under different devices and scanning parameters on subsequent processing.
[0021]
[0022] Where I(x,y) is the gray value of the original image at pixel (x,y), I min and I max These are the minimum and maximum gray values of the original image, respectively, and Inorm(x,y) is the normalized gray value.
[0023] Furthermore, step 2 specifically includes:
[0024] (1) Select scale space and set scale set To adapt to the enhancement needs of blood vessels of different sizes;
[0025] (2) Gaussian smoothing is used for moderate noise reduction; for scale The image after Gaussian smoothing:
[0026]
[0027] in It is a two-dimensional Gaussian kernel:
[0028]
[0029] (3) Calculate the Hessian matrix at each scale. The Hessian matrix of the two-dimensional image at point (x,y) is:
[0030]
[0031] , , Using a Gaussian second derivative kernel and image convolution:
[0032]
[0033]
[0034]
[0035] (4) Find the Hessian eigenvalues and eigenvectors:
[0036] Solve the Hessian matrix for each pixel. Two eigenvalues , :
[0037]
[0038] in ;
[0039] (5) Construct the Frangi vesselness function to calculate the response of blood vessels:
[0040]
[0041] in , , c are preset parameters;
[0042] (6) Multi-scale fusion:
[0043] The final enhanced vascular image is generated by maximizing the vesselness across all scales.
[0044] .
[0045] Furthermore, step 3 specifically includes:
[0046] (1) Set a local window ;
[0047] (2) For each pixel (x,y) in the image, calculate the local mean within its neighborhood window. ;
[0048]
[0049] in, This represents a local window centered at (x, y), with a window size of [value missing]. N is the total number of pixels in the window;
[0050] (3) For each pixel (x,y) in the image, calculate the local standard deviation within its neighborhood window. ;
[0051]
[0052] (4) Calculate the local threshold T(x,y);
[0053]
[0054] Here, parameter k controls the sensitivity of the threshold to local contrast; parameter R is a normalization constant of the standard deviation, used to control the scale of the formula.
[0055] (5) Binarization: For each pixel, compare its grayscale value with the local threshold.
[0056] .
[0057] Furthermore, in step 4, the region is divided into multiple regions of interest (ROIs) based on the ETDRS region division standard. Specifically, the ETDRS grid centered on the fovea is used to locate the ROIs. The avascular area of the fovea is located within a 1mm diameter ring, and the parafoveal ROIs are located between the inner and outer rings. The region is further divided into four quadrants: superior, temporal, inferior, and nasal.
[0058] Furthermore, in step 5, the vascular characteristic parameters include fractal dimension, vascular density, skeletal density, and vascular diameter index; wherein:
[0059] The fractal dimension calculation process is as follows:
[0060] Assuming the size of the two-dimensional grayscale image of the fundus is The grayscale range is [0, ;
[0061] Mapping a two-dimensional grayscale image to a three-dimensional space, with planar coordinates (x, y) and height equal to the grayscale value z, means that the image corresponds to a three-dimensional discrete surface.
[0062] Take the network scale as Where s is the side length of each small piece on the plane, satisfying: ;
[0063] The image plane is divided into several small blocks, each block being [size missing]. The number of blocks is ;
[0064] Let the grayscale axis layer height be:
[0065]
[0066] For the (i,j)th planar grid block, calculate the minimum and maximum values of its pixel grayscale: , ;
[0067] The estimated number of boxes occupied by this planar grid block is: , ;
[0068] If we directly use grayscale values and floor height h, it can be written as: ;
[0069] Summing over all planar mesh blocks yields the total number of boxes at scale s: ;
[0070] If an image has fractal properties, then it satisfies the power law relationship: ;
[0071] Take the logarithm of both sides: ;
[0072] Different sizes of ( , Plot the graph and fit the line using the least squares method; the slope of the line is the estimated value of the fractal dimension.
[0073] The vascular density represents the ratio of the area occupied by blood vessels to the total area of the selected region; the formula for calculating vascular density is:
[0074]
[0075] In the binarized image, the pixel value occupied by blood vessels is 1, which is white;
[0076] The skeleton density represents the length of blood vessels in the skeletonized fundus image. The skeletonized image is generated by reducing the pixels in the width direction of the blood vessels to only 1 pixel. The formula for calculating the skeleton density is:
[0077] ;
[0078] The vessel diameter index represents the actual blood supply to the retina by assessing the average vessel volume, which is calculated using binarized fundus images and skeletonized images.
[0079] .
[0080] Furthermore, in step 5, the texture feature parameters include contrast, correlation, entropy, energy, and homogeneity; wherein:
[0081] The formula for calculating the contrast ratio is as follows:
[0082] Contrast =
[0083] in, This represents the normalized frequency of occurrence of pixel pairs with gray levels i and j under direction θ and distance d.
[0084] The correlation calculation formula is as follows:
[0085] Correlation =
[0086] Where μx and μy are the mean values of the gray-level co-occurrence matrix in the row and column directions, respectively;
[0087]
[0088]
[0089] σx and σy are the corresponding standard deviations:
[0090]
[0091] ;
[0092] The formula for calculating the entropy is as follows:
[0093] Entropy = ;
[0094] The energy calculation formula is as follows:
[0095] Energy = ;
[0096] The formula for calculating homogeneity is as follows:
[0097] Homogeneity = .
[0098] Furthermore, in step 5, the six layered images correspond to the nerve fiber layer vascular plexus, superficial vascular plexus, intermediate capillary plexus, deep capillary plexus, superficial vascular complex, and deep vascular complex, respectively.
[0099] On the other hand, the present invention also provides a device for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging. The device is used to implement the above-described method and includes:
[0100] Optical coherence tomography module: used to acquire tomographic image data of the target area of the fundus;
[0101] Scan control module: Used to control the scanning parameters and scanning path of the optical coherence tomography (OCT) module;
[0102] Eye positioning and fixation guidance module: used to guide the subject to fixate and locate the target eye position;
[0103] Signal acquisition and conversion module: used to acquire OCT interference signals and convert them into digital image data;
[0104] The vessel recognition and extraction module is used to segment and extract the structure of vascular regions in tomographic image data.
[0105] Region segmentation module: used to divide fundus images into multiple regions of interest according to the ETDRS standard;
[0106] Feature parameter calculation module: used to calculate retinal vascular feature parameters and texture feature parameters for each region of interest;
[0107] 3D Reconstruction and Analysis Module: Used to construct a spatial structural model of the fundus vascular network;
[0108] Data storage and display module: used to store and display processing results;
[0109] Main control processing module: Used to coordinate and control the operation of the above modules.
[0110] The present invention has the following advantages over the prior art:
[0111] (1) Improved accuracy and stability of vascular and texture feature parameter extraction. This invention employs a combination of a Hessian matrix-based Frangi vascular filter and a Sauvola thresholding strategy to process OCTA images. Studies have confirmed that this combination achieves the lowest coefficient of variation (CV) in vascular density calculation, indicating the highest measurement stability across different individuals. Furthermore, compared to manually segmented OCTA images, the VD obtained using the Frangi-Sauvola combination has the lowest relative error, validating the effectiveness of this technique.
[0112] (2) The invention achieves the joint extraction of multiple feature parameters. In addition to calculating the vessel density (VD), the invention also extracts multiple vascular feature parameters such as fractal dimension (FD), skeleton density (SKD), and vessel diameter index (VDI). Furthermore, it calculates five complementary texture parameters: contrast, correlation, entropy, energy, and homogeneity, to comprehensively characterize the edge strength, directional regularity, complexity, orderliness, and smoothness of the retinal vascular network. This provides multi-dimensional quantitative indicators for the comprehensive evaluation of the retinal microvascular network.
[0113] (3) Supports comprehensive analysis of multi-layered vascular networks. This invention extracts characteristic parameters from six independent layers: the neurofibrillary plexus (NFLVP), superficial vascular plexus (SVP), intermediate capillary plexus (ICP), deep capillary plexus (DCP), superficial vascular complex (SVC), and deep vascular complex (DVC), which can comprehensively reflect the pathological changes of vascular networks at different depths of the retina. The results show that when distinguishing between patients with mild non-proliferative diabetic retinopathy (NPDR) and normal healthy subjects, the AUROC values of FD for SVP-I and ICP-I reached 0.865 and 0.923, respectively. Among them, the sensitivity of FD for ICP-I was 84.6% and the specificity was 95.0%; the AUROC value of VD for ICP-T reached 0.935, with a sensitivity of 96.2% and a specificity of 90.0%; and the AUROC values of SKD for SVP-I and ICP-I reached 0.95 and 0.944, respectively. Entropy showed significant inter-group differences in multiple quadrants of SVP, ICP, DCP, SVC, and DVC, with diagnostic sensitivity and specificity exceeding 85%. This multi-level analysis capability provides clinicians with a powerful tool for locating the depth distribution and spatial location of early microvascular lesions, whereas most existing technologies are limited to single vascular layers or global analysis.
[0114] (4) It has good cross-device applicability. Research and verification show that the OCTA images processed by the Frangi-Sauvola combination can achieve VD values close to those of manual segmentation on images acquired by different OCTA devices (Spectralis OCTA and RTVue-XR Avanti OCTA system), indicating that the present invention has good cross-device versatility.
[0115] (5) No need for a large amount of labeled data, and high computational efficiency. Compared with deep learning-based methods, this invention uses traditional image processing methods, which do not require a large amount of high-quality manually labeled data for model training. The algorithm is simple to implement, has low computational complexity, and strong interpretability of results, making it suitable for clinical application scenarios with limited computing resources. Attached Figure Description
[0116] Figure 1 This is a flowchart of the vascular feature parameter extraction process of the present invention.
[0117] Figure 2 This is a flowchart of the texture feature parameter extraction process of the present invention.
[0118] Figure 3 This is a block diagram of the device module of the present invention.
[0119] Figure 4 This is a schematic diagram of the region division of the OCTA image in an embodiment of the present invention (OS: left eye; OD: right eye), which shows the ETDRS grid centered on the fovea, including an inner ring with a diameter of 1 mm and an outer ring with a diameter of 3 mm, as well as four quadrants: superior (S), temporal (T), inferior (I), and nasal (N).
[0120] Figure 5 This is a comparison chart of the coefficient of variation of blood vessel density in different retinal regions under different combinations of vascular enhancement techniques and threshold strategies in embodiments of the present invention.
[0121] Figure 6 This is a comparison chart of the relative errors between blood vessel density calculated by different combinations of filters and threshold strategies in embodiments of the present invention and manually labeled data. Detailed Implementation
[0122] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0123] Example 1
[0124] like Figure 1 and 2 As shown, a method for extracting retinal vascular feature parameters and texture feature parameters based on optical coherence tomography (OCT) angiography includes the following steps:
[0125] Step 1: OCTA Image Acquisition and Normalization Preprocessing
[0126] The system receives OCTA images of the fundus and performs normalization preprocessing on the received fundus images to obtain preprocessed images. Normalization maps the gray values of the original image to the [0,1] interval to eliminate the influence of differences in image gray range under different devices and scanning parameters on subsequent processing.
[0127]
[0128] Where I(x,y) is the gray value of the original image at pixel (x,y), I min and I max I represents the minimum and maximum grayscale values of the original image, respectively. norm (x,y) represents the normalized grayscale value.
[0129] Step 2: This step uses the Frangi vessel filter based on the Hessian matrix to enhance the blood vessels in the normalized OCTA image. The Frangi vessel filter was chosen because it uses the eigenvectors of the Hessian matrix to calculate the probability of blood vessels in the image region, effectively enhancing tubular structures and suppressing non-vascular signals. Results show that combining the Frangi vessel filter with the Sauvola thresholding strategy yields the smallest coefficient of variation in blood vessel density calculation and the smallest relative error compared to manual segmentation results.
[0130] (1) Select scale space and set scale set To adapt to the enhancement needs of blood vessels of different sizes;
[0131] (2) Gaussian smoothing is used for moderate noise reduction;
[0132] For scale The image after Gaussian smoothing:
[0133]
[0134] in It is a two-dimensional Gaussian kernel:
[0135]
[0136] (3) Calculate the Hessian matrix at each scale. The Hessian matrix of the two-dimensional image at point (x,y) is:
[0137]
[0138] , , Using a Gaussian second derivative kernel and image convolution:
[0139]
[0140]
[0141]
[0142] (4) Find the Hessian eigenvalues and eigenvectors
[0143] Solve the Hessian matrix for each pixel. Two eigenvalues ,
[0144]
[0145] in ;
[0146] (5) Construct the Frangi vesselness function to calculate the response of blood vessels:
[0147]
[0148] in , , c are preset parameters;
[0149] (6) Multi-scale fusion
[0150] The maximum value of vesselness across all scales is used to generate the final enhanced vascular image;
[0151] .
[0152] Step 3: Perform binarization using a threshold strategy.
[0153] This step employs the Sauvola local thresholding strategy to binarize the enhanced vessel image. The Sauvola thresholding strategy was chosen because it is an adaptive local thresholding technique that improves the Niblack method by introducing a dynamic term based on local contrast and mean, making it particularly suitable for images with uneven backgrounds. The results show that combining the Sauvola thresholding strategy with the Frangi vessel filter yields optimal vessel density calculation results.
[0154] Specifically, it includes the following sub-steps:
[0155] (1) Set a local window ;
[0156] (2) For each pixel (x,y) in the image, calculate the local mean within its neighborhood window. ;
[0157]
[0158] in, This represents a local window centered at (x, y), with a window size of [value missing]. N is the total number of pixels in the window;
[0159] (3) For each pixel (x,y) in the image, calculate the local standard deviation within its neighborhood window. ;
[0160]
[0161] (4) Calculate the local threshold T(x,y):
[0162]
[0163] Here, parameter k controls the sensitivity of the threshold to local contrast; parameter R is a normalization constant of the standard deviation, used to control the scale of the formula.
[0164] (5) Binarization: Compare the gray value of each pixel with the local threshold;
[0165] .
[0166] Step 4: Image region segmentation after binarization
[0167] Based on the size of the received fundus OCTA images and in conjunction with the ETDRS (Early Treatment of Diabetic Retinopathy Study) region division criteria, the region was divided into multiple regions of interest (ROIs). Specifically, an ETDRS grid centered on the fovea was used to locate the ROIs. The foveal avascular zone (FAZ) was located within a 1 mm diameter ring, and the parafoveal ROIs were located between the inner ring (1 mm) and the outer ring (3 mm), and were further divided into four quadrants: Superior (S), Temporal (T), Inferior (I), and Nasal (N).
[0168] Step 5: Calculation of retinal vascular and texture feature parameters
[0169] After obtaining the binarized fundus image through the aforementioned steps, retinal vascular features and texture feature parameters are calculated for each region.
[0170] Retinal vascular features include fractal dimension (FD), vessel density (VD), skeleton density (SKD), and vessel diameter index (VDI).
[0171] (1) Fractal dimension (FD)
[0172] FD represents the irregular and complex pattern of the fundus vascular network. This study used the differential box counting method.
[0173] Assuming the size of the two-dimensional grayscale image of the fundus is The grayscale range is [0, ;
[0174] Mapping a two-dimensional grayscale image to a three-dimensional space, with planar coordinates (x, y) and height equal to the grayscale value z, means that the image corresponds to a three-dimensional discrete surface.
[0175] Take the network scale as Where s is the side length of each small piece on the plane, satisfying: In this way, the image plane can be divided into several small blocks, each block being [size missing]. The number of blocks is ;
[0176] To make the 3D box as close to a cube as possible, the height direction is also divided accordingly. Let the grayscale axis layer height be:
[0177]
[0178] For the (i,j)th planar grid block, calculate the minimum and maximum values of its pixel grayscale: , ;
[0179] If they fall on the k-th and l-th layers of the grayscale axis respectively, then the estimated number of boxes occupied by this block is: , ;
[0180] If we directly use grayscale values and floor height h, it is often written as: ;
[0181] Summing over all planar mesh blocks yields the total number of boxes at scale s: ;
[0182] If an image has fractal properties, then it satisfies the power law relationship: ;
[0183] Take the logarithm of both sides: ;
[0184] Different sizes of ( , Plot the graph and fit the line using the least squares method; the slope of the line is the estimated fractal dimension.
[0185] (2) Vascular density (VD)
[0186] VD represents the ratio of the area occupied by blood vessels to the total area of the selected region. In a binarized image, pixels occupied by blood vessels have a gray value of 1 (white), while pixels not occupied by blood vessels have a gray value of 0 (black). Therefore, the formula for calculating VD is: .
[0187] (3) Skeletal density (SKD)
[0188] SKD represents the length of blood vessels in a skeletonized fundus image. Skeletonized images are generated by reducing the pixels along the width of the blood vessels to only one pixel. The formula for calculating SKD is: .
[0189] (4) Vessel Diameter Index (VDI)
[0190] VDI represents the actual blood supply to the retina by assessing the average vessel volume, which is calculated using binarized fundus images and skeletonized images:
[0191] .
[0192] Retinal texture features include contrast, correlation, entropy, energy, and homogeneity.
[0193] (1) Contrast
[0194] Contrast measures the local variations presented in an image, reflecting the sharpness of textures and edge strength, and highlighting the edges of specific biological structures.
[0195] The formula for calculating contrast is as follows:
[0196] Contrast =
[0197] in, This represents the normalized frequency of pixel pairs of gray levels i and j under direction θ and distance d.
[0198] (2) Correlation
[0199] Correlation measures the linear dependence of gray levels along the row and column directions, reflecting the directionality and structural consistency of the texture. The correlation calculation formula is as follows:
[0200] Correlation =
[0201] Where μx and μy are the mean values of the gray-level co-occurrence matrix in the row and column directions, respectively.
[0202]
[0203]
[0204] σx and σy are the corresponding standard deviations:
[0205]
[0206] .
[0207] (3) Entropy
[0208] Entropy is a measure of information content, reflecting the complexity and disorder of image texture. The calculation formula is as follows:
[0209] Entropy = .
[0210] (4) Energy
[0211] Energy (also known as the second moment of an angle) measures the uniformity or orderliness of image texture. The formula for calculating energy is as follows:
[0212] Energy = .
[0213] (5) Homogeneity
[0214] Homogeneity measures the local uniformity of gray-level distribution in an image, reflecting the smoothness of texture. The formula for calculating homogeneity is as follows:
[0215] Homogeneity = .
[0216] Step Six: Extraction of Feature Parameters from Multilayer Vascular Networks
[0217] Repeat steps 1 to 5 above to extract vascular and texture features from the six layered images of the retina. These six layered images correspond to the nerve fiber layer vascular plexus (NFLVP), superficial vascular plexus (SVP), intermediate capillary plexus (ICP), deep capillary plexus (DCP), superficial vascular complex (SVC), and deep vascular complex (DVC), respectively. The SVC is composed of NFLVP and SVP, and the DVC is composed of ICP and DCP.
[0218] Example 2
[0219] like Figure 2 A device for extracting retinal vascular feature parameters and texture feature parameters based on optical coherence tomography (OCT) vascular imaging, comprising:
[0220] Optical coherence tomography module: used to acquire tomographic image data of the target area of the fundus;
[0221] Scan control module: Used to control the scanning parameters and scanning path of the optical coherence tomography (OCT) module;
[0222] Eye positioning and fixation guidance module: used to guide the subject to fixate and locate the target eye position;
[0223] Signal acquisition and conversion module: used to acquire OCT interference signals and convert them into digital image data;
[0224] The vessel recognition and extraction module is used to segment and extract the structure of the vascular region in the tomographic image data. Specifically, it includes a vessel enhancement unit (using a Frangi vessel filter based on the Hessian matrix for vessel enhancement) and a binarization processing unit (using a Sauvola thresholding strategy for binarization processing).
[0225] Region segmentation module: used to divide fundus images into multiple regions of interest according to the ETDRS standard;
[0226] Feature parameter calculation module: used to calculate the vascular feature parameters and texture feature parameters of each region of interest;
[0227] 3D Reconstruction and Analysis Module: Used to construct a spatial structural model of the fundus vascular network;
[0228] Data storage and display module: used to store and display processing results;
[0229] Main control processing module: Used to coordinate and control the operation of the above modules.
[0230] Example 3
[0231] This embodiment verifies the optimized combination of vascular enhancement technology and thresholding strategy, illustrating that the combination of the Hessian matrix-based Frangi vascular filter and the Sauvola thresholding strategy can achieve the best output effect in vascular density calculation.
[0232] (I) Experimental subjects and data collection
[0233] The right eyes of 19 healthy individuals with a mean age of 42.1 ± 12.6 years were selected. OCTA imaging was performed using a Spectralis OCTA system (Heidelberg Engineering GmbH) with a light source of 870 nm wavelength and 50 nm bandwidth, and a scanning speed of 85,000 scans per second. Six layered images (NFLVP, SVP, ICP, DCP, SVC, and DVC) were generated in PNG format, with an image size of 3 × 3 mm, centered at the fovea.
[0234] (II) Image Processing Methods
[0235] Five vascular enhancement techniques (VET) were used for evaluation:
[0236] J1: No filtering (control group);
[0237] J2: Frangi blood vessel filter based on Hessian matrix;
[0238] J3: Minimum filter;
[0239] J4: Max filter;
[0240] J5: Gabor filter;
[0241] J6: Optimally oriented flux (OOF).
[0242] Six threshold strategies (TS) were used for evaluation:
[0243] K1: Global Otsu;
[0244] K2: Local Otsu;
[0245] K3: Global mean;
[0246] K4: Iterative Self-Organizing Data Analysis Algorithm (ISODATA);
[0247] K5: Niblack;
[0248] K6: Sauvola.
[0249] The OCTA images are binarized by different combinations of the above-mentioned filter and threshold strategies, and the blood vessel density is calculated based on the binarized OCTA images.
[0250] (III) Data Analysis Methods
[0251] Bland-Altman (BA) analysis was used to evaluate the consistency of VD calculated by different combinations of filters and threshold strategies; the optimal method for calculating vessel density was screened by calculating the coefficient of variation (CV).
[0252] In the BA analysis, two comparisons were made: (1) under threshold strategy K1, the VD difference between J2~J6 and J1 was compared; (2) under no filtering condition J1, the VD difference between K2~K6 and K1 was compared.
[0253] (iv) Experimental Results
[0254] Most BA analysis results were statistically significant (p<0.001), indicating significant differences in VD, and these differences depend on the choice of filter or threshold strategy. The results show that CV is minimized when Sauvola (K6) is used in combination with Frangi (J2) or OOF (J6).
[0255] Furthermore, the VD values were calculated using OCTA images exported from another device (RTVue-XR Avanti OCTA system) and compared with the VD values calculated using the corresponding manually segmented images for verification. The results show that the VD obtained from OCTA images processed by Frangi (J2) and Sauvola (K6) is closer to the VD obtained from manually segmented images, with the smallest relative error.
[0256] Example 4
[0257] This embodiment illustrates the application of vascular feature parameters in the classification of fundus diseases; it demonstrates that vascular feature parameters extracted from OCTA images processed by combining Frangi (J2) and Sauvola (K6) can be used for image classification and disease-aided diagnosis.
[0258] (a) Research Subjects
[0259] This study included 26 eyes from 26 healthy subjects, 20 eyes from 20 patients without diabetic retinopathy (NDR), and 20 eyes from 20 patients with mild nonproliferative diabetic retinopathy (NPDR). The mean age of the healthy subjects was 54.2 ± 7.9 years, the mean age of the NDR patients was 54.3 ± 9.1 years, and the mean age of the mild NPDR patients was 58.4 ± 9.6 years.
[0260] The inclusion criteria for diabetic patients are as follows: (1) type 2 diabetes; (2) mild non-proliferative diabetic retinopathy (NPDR) as determined by three experienced ophthalmologists according to the Early Diabetic Retinopathy Treatment Study (ETDRS) criteria, which is defined as the presence of ≥1 microaneurysm without other abnormalities.
[0261] Exclusion criteria are as follows: (1) Subjects with other eye diseases, such as glaucoma, vitreous macular disease, macular edema and age-related macular degeneration; (2) Subjects who have previously received treatment for diabetic retinopathy (DR); (3) Subjects who have previously undergone any ophthalmic surgery; (4) Subjects with poor OCTA image quality and a quality index (Q) below 25.
[0262] (II) Image Processing and Feature Extraction
[0263] OCTA imaging was performed using the Spectralis OCTA system, generating six layered images: NFLVP, SVP, ICP, DCP, SVC, and DVC. Regions of interest (ROIs) were located using an ETDRS grid centered on the fovea, dividing the parafoveal region into four quadrants: superior (S), temporal (T), inferior (I), and nasal (N).
[0264] OCTA images were processed using a combination of a Hessian matrix-based Frangi vessel enhancement filter and a Sauvola thresholding strategy. Four vascular feature parameters—FD, VD, SKD, and VDI—were extracted from the binarized OCTA images.
[0265] (III) Statistical Analysis Methods
[0266] One-way multivariate analysis of variance (MANOVA) and Tukey's post-hoc test were used to compare vascular-related characteristics among the study groups. ROC analysis was used to assess the diagnostic ability of each vascular-related characteristic.
[0267] (iv) Experimental Results
[0268] Fractal Dimension (FD): Compared with normal healthy subjects, NDR patients showed statistically significant differences in FD for SVP-S, SVP-I, and ICP-S, ICP-I; mild NPDR patients showed statistically significant differences in FD for NFLVP-I, NFLVP-T, SVP-S, SVP-I, SVP-T, ICP-S, ICP-I, ICP-T, ICP-N, DCP-I, SVC-S, SVC-I, SVC-T, and DVC-S, DVC-I. When distinguishing between mild NPDR patients and normal healthy subjects, the AUROC values for FD of SVP-I and ICP-I were 0.865 and 0.923, respectively, with ICP-I FD showing a sensitivity of 84.6% and a specificity of 95.0%.
[0269] Vascular density (VD): Compared with normal healthy subjects, NDR patients showed statistically significant differences in VD across multiple quadrants of SVP, ICP, DCP, SVC, and DVC; mild NPDR patients showed statistically significant differences in VD across all regions of the four quadrants of NFLVP-I, NFLVP-T, and SVP, ICP, DCP, SVC, and DVC. In distinguishing mild NPDR patients from normal healthy subjects, the AUROC values for VD in SVP-I and ICP-T reached 0.935 and 0.935, respectively, with ICP-T showing a sensitivity of 96.2% and a specificity of 90.0%.
[0270] Skeletal density (SKD): Compared with normal healthy subjects, NDR patients showed statistically significant differences in SKD across multiple quadrants of SVP, ICP, DCP, SVC, and DVC; mild NPDR patients showed statistically significant differences in SKD across NFLVP-T and all regions of all four quadrants of SVP, ICP, DCP, and DVC, as well as most quadrants of SVC. In distinguishing mild NPDR patients from normal healthy subjects, the AUROC values for SKD of SVP-I and ICP-I were 0.95 and 0.944, respectively, while the SKD values of ICP-I and ICP-T both achieved a sensitivity of 96.2% and a specificity of 90.0%.
[0271] Vascular Diameter Index (VDI): Compared with normal healthy subjects, NDR patients showed statistically significant differences in VDI at NFLVP-T, NFLVP-N, SVP-S, SVP-I, ICP-S, ICP-I, ICP-T, and DVC-S; mild NPDR patients showed statistically significant differences in VDI at NFLVP-T, SVP-S, SVP-I, SVP-T, and ICP-S, ICP-I, ICP-T.
[0272] The above results indicate that the four vascular characteristic parameters extracted in this invention—FD, VD, SKD, and VDI—all have good diagnostic efficacy in distinguishing between normal healthy subjects, NDR patients, and mild NPDR patients, and can serve as effective quantitative indicators for early screening and auxiliary diagnosis of fundus diseases.
[0273] Example 5
[0274] This embodiment verifies the OCTA image preprocessing method based on the Frangi blood vessel filter and the Sauvola thresholding strategy, illustrating the implementation method and its technical effects of using a combination of the Hessian matrix-based Frangi blood vessel filter and the Sauvola local thresholding strategy for OCTA image preprocessing.
[0275] 1. Experimental subjects and data collection
[0276] Eyes from 26 healthy subjects, 20 eyes from 20 patients without diabetic retinopathy (NDR), and 20 eyes from 20 patients with mild nonproliferative diabetic retinopathy (NPDR) were selected. The mean age of the healthy subjects was 54.2 ± 7.9 years, the mean age of the NDR patients was 54.3 ± 9.1 years, and the mean age of the mild NPDR patients was 58.4 ± 9.6 years. There were no statistically significant differences in age among the three groups.
[0277] OCTA imaging was performed using a Spectralis OCTA system (Heidelberg Engineering GmbH, Germany), employing a light source with a wavelength of 870 nm and a bandwidth of 50 nm, and a scanning speed of 85,000 A scans per second. The OCTA algorithm used a full-spectrum probabilistic method to generate 3×3 mm images centered on the fovea. Six layered images were generated, corresponding to the neurofibrillary plexus (NFLVP), superficial vascular plexus (SVP), intermediate capillary plexus (ICP), deep capillary plexus (DCP), superficial vascular complex (SVC), and deep vascular complex (DVC), respectively. The SVC consists of NFLVP and SVP, and the DVC consists of ICP and DCP.
[0278] 2. Image preprocessing methods
[0279] The acquired OCTA images are preprocessed according to the following steps:
[0280] (1) Normalization: Map the gray values of the original OCTA image to the [0,1] interval:
[0281]
[0282] (2) Frangi vessel enhancement based on Hessian matrix: Given a scale set σ∈{0.5,1.0,1.5,2.0,2.5}, calculate the Gaussian smoothed image and Hessian matrix for each scale, solve for the two eigenvalues λ1 and λ2 (|λ1|≤|λ2|) of the Hessian matrix, and construct the Frangi vesselness function:
[0283]
[0284] in , β=0.5, c=0.5max(S). The maximum value of the vesselness response at all scales is taken to form the final enhanced vascular image.
[0285] (3) Sauvola Local Threshold Binarization: Set the local window size to 25×25, calculate the local mean m(x,y) and local standard deviation s(x,y) for each pixel, and calculate the local threshold:
[0286]
[0287] Where k=0.3 and R=0.5. Binarization is performed by comparing the pixel grayscale value with a local threshold.
[0288] 3. Verification of technical effectiveness
[0289] The OCTA images preprocessed as described above show clear vascular structures, and non-vascular signals and noise are effectively suppressed. Compared with images directly based on the original grayscale images, the binarized images preprocessed with Frangi+Sauvola can more accurately reflect the true topological structure of the vascular network, providing a high-quality image foundation for the accurate extraction of subsequent texture features.
[0290] Example 6
[0291] This embodiment is a specific implementation method for extracting texture features from OCTA images, illustrating the specific implementation method for extracting five texture features—contrast, correlation, entropy, energy, and homogeneity—from preprocessed OCTA images.
[0292] 1. Image region segmentation
[0293] The image was divided into regions using an ETDRS (Early Treatment of Diabetic Retinopathy Study) grid centered on the fovea. The parafoveal region (located between the inner ring with a diameter of 1 mm and the outer ring with a diameter of 3 mm) was further divided into four quadrants: Superior (S), Temporal (T), Inferior (I), and Nasal (N), for a total of four regions of interest (ROIs).
[0294] 2. Construction of Gray-Level Co-occurrence Matrix (GLCM)
[0295] For a binary OCTA image, the number of gray levels L=2 (0 represents background, 1 represents blood vessels). For each ROI, the GLCM is calculated at a distance d=1 pixel in four directions (0°, 45°, 90°, 135°) to obtain the normalized joint probability density sθ(i,j|d), where i,j∈{0,1}. The average GLCM of the four directions is taken as the final GLCM.
[0296] 3. Texture Feature Calculation
[0297] Based on GLCM, five texture features are calculated using the following formulas: Contrast, Correlation, Entropy, Energy, and Homogeneity.
[0298] 4. Extraction of texture features from multilayer vascular networks
[0299] Repeat steps 1 to 3 above to extract five texture features for each ROI of the six layered images: NFLVP, SVP, ICP, DCP, SVC, and DVC, resulting in a total of 6 layers × 4 quadrants × 5 features = 120 texture feature values.
[0300] Example 7
[0301] This embodiment illustrates the application of texture features in the diagnosis of diabetic retinopathy, demonstrating the effectiveness of the OCTA image texture features extracted in this invention in distinguishing between normal healthy subjects, NDR patients, and mild NPDR patients.
[0302] 1. Research Subjects and Image Acquisition
[0303] The study subjects were the same as in Example 5. All subjects underwent slit-lamp examination, fundus photography, OCT, and OCTA examinations.
[0304] 2. Texture Feature Extraction
[0305] Following the method described in Example 1, all OCTA images were preprocessed and texture features were extracted.
[0306] 3. Statistical Analysis Methods
[0307] One-way MANOVA and Tukey's post-hoc test were used to compare the differences in texture features among the groups. Receiver operating characteristic (ROC) curve analysis was used to evaluate the diagnostic efficacy of each texture feature, and the area under the curve (AUROC), sensitivity, and specificity were calculated.
[0308] 4. Experimental Results
[0309] (1) Diagnostic efficacy of entropy
[0310] Entropy showed significant inter-group differences and good diagnostic efficacy in multiple quadrants of SVP, ICP, DCP, SVC, and DVC.
[0311] Differentiating NDR from normal eyes: Entropy showed high diagnostic efficacy (AUROC > 0.8) in SVP-I (AUROC = 0.827), SVP-T (AUROC = 0.855), ICP-I (AUROC = 0.815), ICP-T (AUROC = 0.854), DCP-I (AUROC = 0.883), DVC-I (AUROC = 0.860), and DVC-T (AUROC = 0.804). Among them, DCP-I had the highest entropy diagnostic efficacy, with an AUROC of 0.883, a sensitivity of 88.5%, and a specificity of 80.0%.
[0312] Differentiating between mild NPDR and normal eyes: Entropy showed high diagnostic efficacy in all four quadrants of SVP (S: AUROC=0.852, I: AUROC=0.935, T: AUROC=0.933, N: AUROC=0.863), all four quadrants of ICP (S: AUROC=0.879, I: AUROC=0.935, T: AUROC=0.919, N: AUROC=0.871), DCP-I (AUROC=0.921), DCP-T (AUROC=0.883), and all quadrants of DVC. DVC-I showed the highest entropy diagnostic efficacy, with an AUROC of 0.937, a sensitivity of 96.2%, and a specificity of 85.0%.
[0313] Differentiating between mild NPDR and NDR: Entropy showed good diagnostic efficacy (AUROC > 0.7) in SVP-I (AUROC=0.748), ICP-I (AUROC=0.762), ICP-T (AUROC=0.730), SVC-I (AUROC=0.745), and DVC-I (AUROC=0.741). Among them, ICP-I had an entropy AUROC of 0.762, a sensitivity of 65.0%, and a specificity of 85.0%.
[0314] (2) Diagnostic efficacy of correlation
[0315] The correlation coefficient (AUROC) reached 0.802 when DCP-I distinguished between NDR and normal eyes. When distinguishing between mild NPDR and normal eyes, the correlation coefficient (AUROC > 0.8) showed high diagnostic power in SVP-T (AUROC = 0.804) and SVP-N (AUROC = 0.800), as well as in DCP-I (AUROC = 0.852), DCP-T (AUROC = 0.863), DCP-N (AUROC = 0.819), DVC-I (AUROC = 0.800), DVC-T (AUROC = 0.833), and DVC-N (AUROC = 0.804).
[0316] (3) Diagnostic efficacy of contrast
[0317] The contrast ratio achieved an AUROC of 0.808 when differentiating between mild NPDR and normal eyes using ICP-I. When differentiating between mild NPDR and NDR, the contrast ratio demonstrated diagnostic efficacy (AUROC > 0.7) in NFLVP-I (AUROC = 0.725), ICP-I (AUROC = 0.727), DCP-I (AUROC = 0.710), and DVC-I (AUROC = 0.706).
[0318] (4) Comparative analysis of texture features and blood vessel density (VD)
[0319] Entropy and VD showed consistency in diagnosis: when differentiating NDR from normal eyes, both showed high diagnostic power in the lower quadrant and temporal region of SVP, ICP, and DVC (AUROC > 0.8); when differentiating mild NPDR from normal eyes, both showed high diagnostic power in multiple quadrants of SVP, ICP, DCP, SVC, and DVC (AUROC > 0.8). Specifically, in differentiating mild NPDR from normal eyes using SVP-I, the AUROC of entropy was 0.935, and the AUROC of VD was 0.935; in differentiating mild NPDR from normal eyes using ICP-T, the AUROC of entropy was 0.919, and the AUROC of VD was 0.935.
[0320] The above results demonstrate that the five texture features extracted in this invention—contrast, correlation, entropy, energy, and homogeneity—all exhibit good diagnostic efficacy in distinguishing between normal healthy subjects, NDR patients, and patients with mild NPDR. Among these, entropy demonstrates the most outstanding diagnostic efficacy and shows good consistency with traditional vascular density, making it an effective quantitative indicator for early screening and auxiliary diagnosis of fundus diseases. Texture features can reflect changes in vascular density and also reveal additional information such as the complexity and structural regularity of vascular networks; the combined use of both provides a more comprehensive assessment of microvascular abnormalities.
[0321] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.
Claims
1. A method for extracting retinal feature parameters based on optical coherence tomography (OCT) angiography, characterized in that, Includes the following steps: Step 1, OCTA Image Acquisition and Normalization Preprocessing: Receive fundus OCTA images and perform normalization preprocessing on the received fundus images to obtain preprocessed images; Step 2: Perform vessel enhancement processing on the normalized OCTA image using the Frangi vessel filter based on the Hessian matrix; Step 3: Binarization using a thresholding strategy: The Sauvola local thresholding strategy is used to binarize the image after vascular enhancement. Step 4: Binarized image region segmentation: Based on the size of the received fundus OCTA image, the region is divided into multiple regions of interest; Step 5: Calculation of retinal feature parameters: After obtaining the binarized fundus image, calculate the retinal vascular feature parameters and texture feature parameters for each region; Step 6: Repeat steps 1 to 5 to extract vascular feature parameters and texture feature parameters from the six layered images of the retina.
2. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging according to claim 1, characterized in that, In step 1, the original image grayscale values are mapped to the [0,1] interval through normalization processing to eliminate the impact of differences in image grayscale range under different devices and scanning parameters on subsequent processing. Where I(x,y) is the gray value of the original image at pixel (x,y), I min and I max These are the minimum and maximum gray values of the original image, respectively, and Inorm(x,y) is the normalized gray value.
3. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging according to claim 1, characterized in that, Step 2 specifically includes: (1) Select scale space and set scale set To adapt to the enhancement needs of blood vessels of different sizes; (2) Gaussian smoothing is used for moderate noise reduction; for scale The image after Gaussian smoothing: in It is a two-dimensional Gaussian kernel: (3) Calculate the Hessian matrix at each scale. The Hessian matrix of the two-dimensional image at point (x,y) is: , , Using a Gaussian second derivative kernel and image convolution: (4) Find the Hessian eigenvalues and eigenvectors: Solve the Hessian matrix for each pixel. Two eigenvalues , : in ; (5) Construct the Frangi vesselness function to calculate the response of blood vessels: in , , c are preset parameters; (6) Multi-scale fusion: The final enhanced vascular image is generated by maximizing the vesselness across all scales. 。 4. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) angiography according to claim 1, characterized in that, Step 3 specifically includes: (1) Set a local window ; (2) For each pixel (x,y) in the image, calculate the local mean within its neighborhood window. ; in, This represents a local window centered at (x, y), with a window size of [value missing]. N is the total number of pixels in the window; (3) For each pixel (x,y) in the image, calculate the local standard deviation within its neighborhood window. ; (4) Calculate the local threshold T(x,y); Here, parameter k controls the sensitivity of the threshold to local contrast; parameter R is a normalization constant of the standard deviation, used to control the scale of the formula. (5) Binarization: For each pixel, compare its grayscale value with the local threshold. 。 5. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging according to claim 1, characterized in that, In step 4, the region is divided into multiple regions of interest (ROIs) based on the ETDRS region division standard. Specifically, the ETDRS grid centered on the fovea is used to locate the ROIs. The avascular area of the fovea is located within a 1mm diameter ring, and the parafoveal ROIs are located between the inner and outer rings. The region is further divided into four quadrants: superior, temporal, inferior, and nasal.
6. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) angiography according to claim 1, characterized in that, In step 5, the vascular characteristic parameters include fractal dimension, vascular density, skeletal density, and vascular diameter index; wherein: The fractal dimension calculation process is as follows: Assuming the size of the two-dimensional grayscale image of the fundus is The grayscale range is [0, ; Mapping a two-dimensional grayscale image to a three-dimensional space, with planar coordinates (x, y) and height equal to the grayscale value z, means that the image corresponds to a three-dimensional discrete surface. Take the network scale as Where s is the side length of each small piece on the plane, satisfying: ; The image plane is divided into several small blocks, each block being [size missing]. The number of blocks is ; Let the grayscale axis layer height be: For the (i,j)th planar grid block, calculate the minimum and maximum values of its pixel grayscale: , ; The estimated number of boxes occupied by this planar grid block is: , ; If we directly use grayscale values and floor height h, it can be written as: ; Summing over all planar mesh blocks yields the total number of boxes at scale s: ; If an image has fractal properties, then it satisfies the power law relationship: ; Take the logarithm of both sides: ; Different sizes of ( , Plot the graph and fit the line using the least squares method; the slope of the line is the estimated value of the fractal dimension. The vascular density represents the ratio of the area occupied by blood vessels to the total area of the selected region; the formula for calculating vascular density is: In the binarized image, the pixel value occupied by blood vessels is 1, which is white; The skeleton density represents the length of blood vessels in the skeletonized fundus image. The skeletonized image is generated by reducing the pixels in the width direction of the blood vessels to only 1 pixel. The formula for calculating the skeleton density is: ; The vessel diameter index represents the actual blood supply to the retina by assessing the average vessel volume, which is calculated using binarized fundus images and skeletonized images. 。 7. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging according to claim 1, characterized in that, In step 5, the texture feature parameters include contrast, correlation, entropy, energy, and homogeneity; wherein: The formula for calculating the contrast ratio is as follows: Contrast = in, This represents the normalized frequency of occurrence of pixel pairs with gray levels i and j under direction θ and distance d. The correlation calculation formula is as follows: Correlation = Where μx and μy are the mean values of the gray-level co-occurrence matrix in the row and column directions, respectively; σx and σy are the corresponding standard deviations: ; The formula for calculating the entropy is as follows: Entropy = ; The energy calculation formula is as follows: Energy = ; The formula for calculating homogeneity is as follows: Homogeneity = 。 8. The method for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging according to claim 1, characterized in that, In step 5, the six layered images correspond to the nerve fiber layer vascular plexus, superficial vascular plexus, intermediate capillary plexus, deep capillary plexus, superficial vascular complex, and deep vascular complex, respectively.
9. A device for extracting retinal feature parameters based on optical coherence tomography (OCT) vascular imaging, the device being used to implement the method according to any one of claims 1 to 8, characterized in that, include: Optical coherence tomography module: used to acquire tomographic image data of the target area of the fundus; Scan control module: Used to control the scanning parameters and scanning path of the optical coherence tomography (OCT) module; Eye positioning and fixation guidance module: used to guide the subject to fixate and locate the target eye position; Signal acquisition and conversion module: used to acquire OCT interference signals and convert them into digital image data; The vessel recognition and extraction module is used to segment and extract the structure of vascular regions in tomographic image data. Region segmentation module: used to divide fundus images into multiple regions of interest according to the ETDRS standard; Feature parameter calculation module: used to calculate retinal vascular feature parameters and texture feature parameters for each region of interest; 3D Reconstruction and Analysis Module: Used to construct a spatial structural model of the fundus vascular network; Data storage and display module: used to store and display processing results; Main control processing module: Used to coordinate and control the operation of the above modules.
