Vascular morphological feature quantification method based on fundus image and gestational hypertension detection device
By using a quantification method based on vascular morphology features from fundus images, the problems of insufficient precision in retinal microvessel segmentation and accuracy in detecting gestational hypertension have been solved, enabling early and accurate detection of gestational hypertension. This method is applicable to fundus cameras with narrow imaging angles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies suffer from insufficient segmentation accuracy in retinal microvessel image analysis, insufficient precision in detecting gestational hypertension, and a lack of real-time processing capabilities and cross-subject adaptability, making it difficult to meet the needs of early diagnosis.
A vascular morphology feature quantification method based on fundus images was adopted. The Otsu adaptive thresholding method and morphological closure operation were used for accurate segmentation. The center line of the blood vessels was extracted by connecting region analysis and skeletonization. Multidimensional features such as arteriovenous diameter ratio, gray intensity and morphological features were used for classification. A Gaussian probability scoring function was constructed to identify arteries and veins. The trained classification model was then used to detect gestational hypertension.
It improves the precision of retinal microvessel segmentation and the accuracy and real-time performance of gestational hypertension detection, is suitable for fundus cameras with narrow imaging angles, reduces the misdiagnosis rate, and improves the reliability and efficiency of clinical diagnosis.
Smart Images
Figure CN121861010A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology and relates to the feature quantification of retinal microvessels, specifically to a method for quantifying vascular morphological features based on fundus images and a device for detecting hypertension during pregnancy. Background Technology
[0002] Retinal microvessels are the only deep microvessels on the human body that can be directly observed non-invasively. Clinically, changes in the contours and characteristics of retinal microvessels are believed to be related not only to ophthalmic diseases but also to systemic cardiovascular diseases. However, current clinical methods for detecting hypertension primarily rely on blood pressure monitoring and urine protein testing, which suffer from delays and insufficient application of fundus imaging. This may be because fundus imaging requires specialized ophthalmologists to capture and analyze the images, demanding high levels of expertise from the operators and consuming significant human resources. Furthermore, early, minute changes are difficult to detect with the naked eye, easily leading to misdiagnosis due to subjective perception or lack of experience.
[0003] By developing repeatable computer programs that provide automated or semi-automated image analysis methods, quantitative measurement and analysis of retinal vessel morphology in fundus images can be performed. This provides auxiliary reference information for clinical staff, thereby improving image reading efficiency, reducing the workload of doctors, and enhancing the reliability and accuracy of disease analysis and diagnosis.
[0004] Accurate segmentation of vascular structures faces significant challenges due to the uneven grayscale distribution, complex vascular structures, low contrast between target vessels and the background, and image noise in fundus images. Traditional segmentation methods often suffer from insufficient accuracy, frequently resulting in vascular rupture and failing to accurately reflect vascular morphological characteristics. While some methods employ deep learning for automated segmentation and feature extraction and analysis, these models typically require substantial training data and time to achieve the desired accuracy and demand high-performance hardware. This may hinder real-time processing in practical applications and lacks cross-subject adaptability. Therefore, current technologies still fall short in areas such as small-sample data adaptation, multi-dimensional feature integration, and clinical applicability.
[0005] Furthermore, existing methods for detecting cardiovascular diseases based on retinal vascular features only consider the arterial-venous diameter ratio, failing to comprehensively utilize information from other morphological features. This results in high requirements for the precision of arterial and venous segmentation and diameter calculation, and frequent misdiagnosis. Moreover, research on the specific population of pregnant women is limited, making it difficult to meet the precise needs for early diagnosis of gestational hypertension.
[0006] In conclusion, quantifying and analyzing the characteristics of retinal microvessels is of great significance in assisting clinicians in the early diagnosis and prevention of related diseases. Summary of the Invention
[0007] To address the shortcomings of existing technologies, this invention proposes a method for quantifying vascular morphological features based on fundus images and a device for detecting gestational hypertension. Based on fundus images, the method performs structural segmentation and morphological feature quantification analysis of retinal vessels, and integrates the arteriovenous diameter ratio and high-level vascular morphological features to assist in the detection of gestational hypertension, thereby improving the reliability of the detection results.
[0008] The method for quantifying vascular morphology features based on fundus images includes the following steps:
[0009] Step 1: Collect fundus images and perform preprocessing and vascular enhancement.
[0010] Step 2: Extract grayscale images from the preprocessed images, perform accurate segmentation using the Otsu adaptive thresholding method, and then remove isolated points using median filtering; eliminate vascular ruptures through morphological closing operations, and remove false detection regions using connected component analysis to achieve high-precision vascular structure extraction.
[0011] Step 3: Use connected component analysis to screen major blood vessel branches, and obtain the topology of the blood vessel centerline through skeleton extraction algorithm; use the Sobel gradient operator to calculate the normal direction of the blood vessel centerline, search for blood vessel boundary points in both directions in the vertical direction, and accurately calculate the local blood vessel diameter based on the Euclidean distance formula.
[0012] Step 4: Based on the precisely segmented vascular structure, extract three vascular diameter features, grayscale intensity features, and morphological features.
[0013] A two-stage classification strategy is used to determine the type of blood vessel. The first stage performs preliminary screening based on a diameter threshold, classifying blood vessels with a diameter feature value less than 6.5 pixels as arteries, blood vessels with a diameter feature value greater than 7.5 pixels as veins, and blood vessels with a diameter feature value in the middle range enter the second stage for fine classification.
[0014] In the second stage of fine classification, a Gaussian probability scoring function is constructed for the extracted features to quantify their correlation with arterial reference features. 1. Reference characteristics of veins The matching degree is determined, and a weighted fusion strategy is used to calculate the total score under the three feature dimensions. :
[0015]
[0016] Where m = artery or vein, representing the index of blood vessel type. Let j be the feature value of the j-th dimension of the blood vessel. For tolerance parameters, The weights are for the corresponding features. The reference feature values for arteries and veins in the j-th dimension are updated based on the preliminary screening results from the first stage. For the blood vessel to be classified, its matching degree S with the arterial reference feature is compared. 动脉 The degree of matching of vein reference features S 静脉 The size of S 静脉 >S 动脉 If the value is +0.05, the vessel is classified as a vein; otherwise, it is classified as an artery. Finally, morphological post-processing is performed on the classification results, using a 3×3 structuring element to perform a closing operation, eliminating small inconsistencies in classification and smoothing the classification boundaries.
[0017] As a preferred option, in the second stage of fine classification, if |S 静脉 -S 动脉 If |≤0.05, a neighboring vessel voting mechanism is introduced. The classification results are counted within a 5×5 neighborhood window of the center point of the vessel to be classified. If the number of arterial pixels in the neighborhood window is greater than the number of vein pixels, the vessel to be classified is determined to be an artery; otherwise, it is determined to be a vein.
[0018] Step 5: Based on the precisely segmented vascular structure and the topology of the vascular centerline, calculate the vascular classification dimension and vascular tortuosity.
[0019] A device for detecting gestational hypertension based on quantified vascular morphology features from fundus images includes:
[0020] The fundus image preprocessing and vascular enhancement module is used to select the green channel from the original fundus image and perform illumination correction and vascular detail enhancement.
[0021] The vessel segmentation module is used to perform thresholding, denoising, and morphological operations on the preprocessed and enhanced binary images of blood vessels to separate the blood vessel part from the background part and obtain a complete and continuous blood vessel segmentation result.
[0022] The blood vessel centerline extraction and diameter measurement module performs vascular skeletonization processing and local diameter calculation on the blood vessel segmentation results, and outputs the average blood vessel diameter and diameter distribution.
[0023] The arteriovenous detection module performs initial screening based on vessel width, and then combines vessel diameter features, grayscale features, and morphological features to construct a Gaussian probability scoring function to quantify the degree of matching between the comprehensive features and the standard features of arteries and veins, and perform fine classification.
[0024] The advanced feature analysis module for vascular morphology calculates the fractal dimension and tortuosity of blood vessels based on the results of blood vessel segmentation and skeletonization.
[0025] The hypertension detection module calculates the average diameter ratio (AVR) of arteries and veins based on the classification results of the arteriovenous detection module. When AVR < 0.7, the AVR value, classification dimension, and vascular tortuosity are normalized to form multidimensional feature parameters. The trained classification model is then used to predict the probability of gestational hypertension and output the predicted probability and the quantitative results of vascular morphology indicators.
[0026] Preferably, the classification model is an SVM, logistic regression, random forest, or KNN algorithm.
[0027] The present invention has the following beneficial effects:
[0028] 1. It takes into account quantitative parameters such as classification dimension, vascular curvature and arteriovenous diameter ratio, making it more suitable for fundus images captured by fundus cameras with narrower imaging angles.
[0029] 2. A two-stage arteriovenous classification method, from coarse to fine, is adopted. Especially for fundus cameras with narrower imaging angles, it can make full use of the detailed information in fundus images, and the classification results are more accurate in areas where arteries and veins are not easy to distinguish.
[0030] 3. Classification methods based on multidimensional parameters are more robust than those based on single feature points alone. The selected multidimensional parameters include the arteriovenous diameter ratio, classification dimension, and vascular tortuosity, which are independent of each other. The fusion of multidimensional features can better improve the accuracy of prediction. Attached Figure Description
[0031] Figure 1 This is a flowchart of a method for quantifying vascular morphological features based on fundus images;
[0032] Figure 2 This is a schematic diagram of fundus image preprocessing and vascular enhancement;
[0033] Figure 3 This is a schematic diagram of the segmentation results of the blood vessel structure;
[0034] Figure 4 This is a schematic diagram of blood vessel centerline extraction and diameter measurement;
[0035] Figure 5 This is a schematic diagram of the results of arteriovenous vascular examination;
[0036] Figure 6 This is a schematic diagram of gestational hypertension detection based on the quantification of vascular morphology features in fundus images.
[0037] Figure 7 It is the result of quantification of vascular morphological characteristics and prediction of gestational hypertension. Detailed Implementation
[0038] The present invention will be further explained below with reference to the accompanying drawings;
[0039] Methods for quantifying vascular morphology features based on fundus images, such as Figure 1 As shown, the specific steps include:
[0040] Step 1: Fundus Image Preprocessing and Vascular Enhancement
[0041] The raw data in this embodiment comes from black and white fundus images acquired by Yuyao Maternal and Child Health Hospital using a Heidelberg retinal angiography system, such as... Figure 2 As shown, two independent optimized top-hat transformations are performed on the input raw fundus image using structuring elements of different scales to suppress background and noise. A large open kernel and a small closed kernel are used to obtain the enhanced coarse blood vessel structure, and a small open kernel and a large closed kernel are used to obtain the enhanced fine blood vessel structure. Then, for the enhanced coarse blood vessel structure, a Gaussian high-pass filter with a spatial scale parameter σ=2 is used for homomorphic filtering to enhance the edges of the coarse blood vessels. For the enhanced fine blood vessel structure, a Gaussian high-pass filter with a spatial scale parameter σ=20 is used for homomorphic filtering to avoid over-enhancing noise. By using different spatial scale parameters to process blood vessel structures of different scales, scale-adaptive contrast enhancement is achieved, avoiding interference from uneven illumination in blood vessel extraction, and further improving the local contrast between blood vessels and the background. Finally, the filtered results are fused, as shown... Figure 2 As shown, the preprocessed fundus image has a clearer vascular structure, which facilitates subsequent segmentation of the vascular structure.
[0042] Step 2: Blood vessel segmentation
[0043] The preprocessed grayscale image was precisely segmented using the Otsu adaptive thresholding method, and false positives were removed using connected component analysis. Median filtering was then used to remove image noise and preserve major blood vessels. Next, `cv2.connectedComponentsWithStats()` was used to find connected components in the image, and regions smaller than `min_size` were filtered out based on their area. Finally, a set of morphology-extended functions, `cv2.morphologyEx()`, with the parameter `cv2.MORPH_CLOSE`, were used to fill small holes in the blood vessels using closing operations, eliminating vessel breaks and ensuring vessel continuity.
[0044] Finally, connected component analysis was performed on the binary image after blood vessel extraction to filter out small regions. Each connected region was assigned a different color, and then the colored blood vessel regions were superimposed onto the original image to achieve high-precision blood vessel structure extraction. The segmentation results are as follows: Figure 3 As shown, the blood vessels are effectively separated from the background, resulting in a binarized vascular structure mask.
[0045] Step 3: Quantitative measurement of blood vessel diameter
[0046] First, the binarized vascular image is labeled with `measure.label()` to identify all independent vascular branches. An area threshold of 500 pixels is set to filter out small, noisy branches, and morphological features are used to exclude non-vascular structures while retaining the main vascular structures. Then, the `morphology.skeletonize()` algorithm is used to extract the vascular centerline with a single-pixel width, establishing a topological skeleton network of the blood vessels. Next, the normal vector of the vascular centerline is calculated using the Sobel gradient operator based on the image gradient field. A bidirectional linear search is then performed along the direction of the normal vector, extending the search to both sides for each sampling point until background pixels are encountered. The inner and outer boundary points of the blood vessel are located and their coordinates are recorded. The Euclidean distance between the inner and outer boundary points is calculated based on these coordinates as the local vascular diameter.
[0047] Finally, statistical analysis was performed on the local blood vessel diameter measurements at all sampling points, outputting the average blood vessel diameter and diameter distribution. The measurement locations and centerline structures were then visualized on the blood vessel images, such as... Figure 4 As shown.
[0048] Step 4: Quantitative Analysis of Arterial and Venous Classification
[0049] Feature extraction is performed based on the topological backbone network of blood vessels established in step three:
[0050] ① Characteristics of blood vessel diameter
[0051] The average diameter D of each blood vessel segment was accurately calculated using distance transformation techniques.
[0052]
[0053] in, Represents the distance transformation value. , ...
[0054] ② Gray intensity characteristics
[0055] Calculate the average brightness value of the blood vessel segment on the original grayscale image. :
[0056]
[0057] in, This represents the grayscale value of the i-th pixel. Because arterial blood has a higher oxygen content, its light reflection characteristics differ from those of veins. Therefore, grayscale intensity features can optically capture the differences between arteries and veins.
[0058] ③ Morphological characteristics
[0059] Calculate the eccentricity and circumscribed rectangle fill degree of each blood vessel segment. The eccentricity describes the straightness or curvature of the blood vessel's course, while the circumscribed rectangle fill degree reflects the compactness of the blood vessel's morphology.
[0060] Based on the feature extraction results, a two-stage classification strategy is implemented. In the first stage, preliminary screening is performed based on the diameter threshold, clearly classifying blood vessels with a diameter feature of less than 6.5 pixels as arteries and blood vessels with a diameter feature of more than 7.5 pixels as veins.
[0061] For vessels with diameters falling within the intermediate range, a second stage of fine-grained classification is performed. A Gaussian probability scoring function is constructed for the extracted features to quantify their correlation with arterial reference features. 1. Reference characteristics of veins The matching degree is determined, and a weighted fusion strategy is used to calculate the total score under the three feature dimensions. :
[0062]
[0063] Where m = artery or vein, representing the index of blood vessel type. Let j be the feature value of the j-th dimension of the blood vessel. This is the tolerance parameter. The reference feature value for arteries or veins in the j-th dimension is updated based on the preliminary screening results of the first stage. This adaptive mechanism ensures that the classification criteria can adapt to the specific feature distribution of the current image, thereby improving the robustness of the algorithm. The weights for corresponding dimensional features reflect the varying importance of different features. In this embodiment, considering that vessel diameter is the most reliable criterion for distinguishing between arteries and veins, the weight of vessel diameter is set to 60%, while the weights of grayscale intensity and morphological features are both 20%. For the vessel to be classified, its matching degree S with the arterial reference feature is compared. 动脉 The degree of matching of vein reference features S 静脉 Size, S 动脉 >S 静脉 If the value is +0.05, the vessel is determined to be an artery. If S 静脉 >S 动脉 If the value is +0.05, then the blood vessel is determined to be a vein. If |S 静脉 -S 动脉If |≤0.05, a neighboring vessel voting mechanism is introduced. Within a 5×5 neighborhood window of the center point of the vessel to be classified, the already classified results are counted. If the number of arterial pixels in the neighborhood window is greater than the number of vein pixels, the vessel to be classified is determined to be an artery; otherwise, it is determined to be a vein. For boundary cases where the vessel diameter feature is between 6.5 and 7.5 pixels, a second-stage fine classification considers its spatial context information to ensure that the classification results of adjacent vessels are biologically reasonable and to avoid isolated misclassifications.
[0064] Morphological post-processing was performed on the classification results, using a 3×3 structuring element to perform a closing operation, eliminating small regions of inconsistency in classification and smoothing classification boundaries. Finally, the classification results were clearly annotated on the original fundus vessel images using visualization methods, such as... Figure 5 As shown, red represents arteries and blue represents veins.
[0065] Step 5: Advanced vascular morphological feature analysis
[0066] Using the Box Counting Method, the number of grid cells containing foreground pixels was counted by covering the image with grids of different sizes. After taking the logarithm of the grid size and the number of grid cells, a linear regression was used to fit the slope to obtain the fractal dimension features of the blood vessels.
[0067] Based on the skeletal vascular centerline, an ordered sequence of center points is first extracted. For each inner point, symmetrical auxiliary points are selected to form a measurement triangle. The lengths of the three sides are calculated using Euclidean distance, and the area of the triangle is determined using Heron's formula, thus deriving the radius of the circumcircle. Calculate curvature Finally, the average curvature and curvature range of the blood vessels are statistically analyzed, and the morphological tortuosity characteristic parameters of the blood vessels are output.
[0068] Step Six: Prediction of Gestational Hypertension
[0069] The probability of gestational hypertension is predicted based on the quantitative results of vascular morphology features in fundus images.
[0070] Under normal conditions, the arteriovenous diameter ratio (AVR) in fundus images is between 0.7 and 0.8. An AVR below 0.7 indicates localized arterial spasm and mild venous dilation, suggesting possible hypertension. To improve prediction accuracy, AVR values, classification dimensions, vascular tortuosity, and patient characteristics are normalized to form 10-dimensional feature parameters, which are then input into a trained classification model to predict the probability of gestational hypertension. The AVR values include the mean AVR across the entire image, the mean AVR in quadrants 1, 2, 3, and 4, and the vascular tortuosity includes the mean tortuosity, maximum tortuosity, and tortuosity variance.
[0071] In this embodiment, SVM is selected as the classification model. The training process of the classification model involves dividing all samples into training data and test data in an 8:2 ratio to train the SVM, using a Gaussian kernel as the kernel function, outputting the probability value of gestational hypertension based on the input data, and using AUC as the evaluation metric to evaluate the predictive performance of the classification model.
[0072] like Figure 6 As shown, the device for detecting gestational hypertension based on the quantification of vascular morphology features from fundus images includes:
[0073] The fundus image preprocessing and vascular enhancement module is used to select the green channel from the original fundus image and perform illumination correction and vascular detail enhancement.
[0074] The vessel segmentation module is used to perform thresholding, denoising, and morphological operations on the preprocessed and enhanced binary images of blood vessels to separate the blood vessel part from the background part and obtain a complete and continuous blood vessel segmentation result.
[0075] The blood vessel centerline extraction and diameter measurement module performs vascular skeletonization processing and local diameter calculation on the blood vessel segmentation results, and outputs the average blood vessel diameter and diameter distribution.
[0076] The arteriovenous detection module performs initial screening based on vessel width, and then combines vessel diameter features, grayscale features, and morphological features to construct a Gaussian probability scoring function to quantify the degree of matching between the comprehensive features and the standard features of arteries and veins, and perform fine classification.
[0077] The advanced feature analysis module for vascular morphology calculates the fractal dimension and tortuosity of blood vessels based on the results of blood vessel segmentation and skeletonization.
[0078] The hypertension detection module calculates the average diameter ratio (AVR) of arteries and veins based on the classification results from the arteriovenous detection module. When AVR ≥ 0.7, the risk level of gestational hypertension is output as low. When AVR < 0.7, the AVR value, classification dimension, and vascular tortuosity are normalized to form multidimensional feature parameters. The trained classification model is then used to predict the probability of gestational hypertension, and the prediction results and quantitative results of vascular morphology indicators are output.
[0079] by Figure 2 Taking the original fundus image shown in the image as an example, after being input into the pregnancy-induced hypertension detection device, it automatically performs feature extraction and quantification, and finally outputs vascular morphology indicators and the probability and level of hypertension risk, such as... Figure 7As shown, the output hypertension risk assessment value can assist doctors in confirming the disease. Simultaneously, the output vascular morphology indicators can provide reference information, highlighting subtle changes in blood vessels, improving clinicians' image interpretation speed, and reducing the probability of missed diagnoses. This method is more accurate than single-feature methods and is better suited to fundus images scanned using narrow-angle confocal laser imaging, exhibiting high sensitivity and specificity in early pregnancy hypertension screening.
Claims
1. A method for quantifying vascular morphological features based on fundus images, characterized in that: The collected fundus images were preprocessed and enhanced with vascular enhancement, then the vascular structure was extracted, and then the diameter, gray intensity and morphological features of the blood vessels were extracted. The blood vessel type was determined by a two-stage classification strategy. The first stage is based on a diameter threshold for preliminary screening. Blood vessels with a diameter feature value less than 6.5 pixels are classified as arteries, and those with a diameter feature value greater than 7.5 pixels are classified as veins. Blood vessels with a diameter feature value in the middle range are then classified into the second stage of fine classification. In the second stage of fine classification, a Gaussian probability scoring function is constructed for the extracted vascular features to calculate the vascular features and arterial reference features.
1. Reference characteristics of veins Match score : Where m = artery or vein, representing the index of blood vessel type; Let j be the feature value of the j-th dimension of the blood vessel. For tolerance parameters, The weights are for the corresponding features; Let S be the reference feature value of the artery or vein in the j-th dimension; for the blood vessel to be classified, if S... 静脉 >S 动脉 If the value is +0.05, the vessel is determined to be a vein; otherwise, it is determined to be an artery. The topological structure of the blood vessel centerline is obtained by using a skeletonization extraction algorithm, and the classification dimension and tortuosity of the blood vessel are calculated by combining the blood vessel structure.
2. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: Two independent optimized top-hat transformations are performed on the input raw fundus image using structuring elements of different scales. The enhanced coarse blood vessel structure is obtained using a large open kernel and a small closed kernel, and the enhanced fine blood vessel structure is obtained using a small open kernel and a large closed kernel. Then, homomorphic filtering is performed on the enhanced coarse and fine blood vessel structures using Gaussian high-pass filters with different spatial scale parameters to achieve scale-adaptive contrast enhancement. Finally, the filtered results are fused to complete the preprocessing and blood vessel enhancement.
3. The method for quantifying vascular morphological features based on fundus images as described in claim 2, characterized in that: For the enhanced coarse blood vessel structure, a Gaussian high-pass filter with spatial scale parameter σ=2 is used for homomorphic filtering to enhance the edges of the coarse blood vessels; for the enhanced fine blood vessel structure, a Gaussian high-pass filter with spatial scale parameter σ=20 is used for homomorphic filtering to avoid excessive noise enhancement.
4. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: The preprocessed grayscale image is segmented using the Otsu adaptive thresholding method, and false detection regions are removed using connected component analysis. Median filtering is then used to remove image noise. Next, cv2.connectedComponentsWithStats() is used to find connected components in the image, and regions smaller than the minimum area threshold min_size are filtered out based on their area. Then, a set of morphological extension functions, cv2.morphologyEx(), is used to fill small holes in blood vessels through closing operations, eliminating vessel breaks. Finally, connected component analysis is performed on the binary image after blood vessel extraction to filter out small regions, outputting a binarized blood vessel structure mask.
5. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: For the extracted vascular structure, the morphology.skeletonize() algorithm is used to extract the vascular centerline with a single pixel width, and a topological skeleton network of the vascular structure is established. Next, the normal vector of the vascular centerline is calculated based on the image gradient field using the Sobel gradient operator. By using a bidirectional linear search method in the direction of the normal vector, for each sampling point, the search is extended to both sides until the background pixel is encountered. The inner and outer boundary points of the vascular structure are located and their coordinates are recorded. The Euclidean distance between the inner and outer boundary points is calculated based on the coordinates as the local vascular diameter.
6. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: The average diameter D of each blood vessel segment is calculated using distance transformation techniques as a characteristic of the blood vessel's diameter. in, Represents the distance transformation value. , represents the horizontal and vertical coordinates of the i-th pixel, respectively, and N represents the number of pixels in the blood vessel segment; Calculate the average brightness value of the blood vessel segment on the original grayscale image. As a feature of grayscale intensity: in, This represents the grayscale value of the i-th pixel; The eccentricity and circumscribed rectangle filling degree of each blood vessel segment are calculated as morphological features.
7. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: In the second stage of fine classification, if |S 静脉 -S 动脉 If |≤0.05, the classification results are counted within a 5×5 neighborhood window of the center point of the blood vessel to be classified. If the number of arterial pixels is greater than the number of vein pixels, the blood vessel to be classified is determined to be an artery; otherwise, it is determined to be a vein.
8. The method for quantifying vascular morphological features based on fundus images as described in claim 1, characterized in that: Using box counting, the number of grids containing foreground pixels was counted by covering the image with grids of different sizes; after taking the logarithm of the grid size and the number of grids, the slope was fitted by linear regression to obtain the fractal dimension features of blood vessels. Based on the skeletal vascular centerline, an ordered sequence of center points is first extracted. For each inner point, symmetrical auxiliary points are selected to form a measurement triangle. The lengths of the three sides are calculated using Euclidean distance, and the area of the triangle is determined using Heron's formula, thus deriving the radius of the circumcircle. Calculate curvature Finally, the average curvature and curvature range of the blood vessels are statistically analyzed, and the morphological tortuosity characteristic parameters of the blood vessels are output.
9. A device for detecting gestational hypertension based on quantification of vascular morphology features from fundus images, characterized in that: The vascular features in the input fundus image are quantified using the method described in any one of claims 1 to 8 to obtain parameters such as vascular type, vascular diameter, fractal dimension features, and vascular morphology curvature. These parameters, along with the subject's age corresponding to the fundus image, constitute multidimensional feature parameters, which are then input into the trained classification model to obtain the subject's hypertension risk probability.
10. The gestational hypertension detection device based on quantification of vascular morphology features from fundus images as described in claim 9, characterized in that: include: The fundus image preprocessing and vascular enhancement module is used to perform illumination correction and vascular detail enhancement on the original fundus images; The vessel segmentation module is used to perform thresholding, denoising, and morphological operations on the preprocessed and enhanced binary images of blood vessels, separating the blood vessel part from the background part, and obtaining complete and continuous blood vessel segmentation results. The blood vessel centerline extraction and diameter measurement module performs vascular skeletonization and local diameter calculation on the blood vessel segmentation results, and outputs the average blood vessel diameter and diameter distribution. The arteriovenous detection module performs initial screening based on vessel width, and then combines vessel diameter features, grayscale features, and morphological features to construct a Gaussian probability scoring function to quantify the degree of matching between the comprehensive features and the standard features of arteries and veins, and to perform fine classification. The advanced feature analysis module for vascular morphology calculates the fractal dimension and tortuosity of blood vessels based on the results of blood vessel segmentation and skeletonization. The hypertension detection module calculates the average diameter ratio (AVR) of arteries and veins based on the classification results of the arteriovenous detection module. When AVR ≥ 0.7, it is determined that there is no gestational hypertension. When AVR < 0.7, the AVR value, classification dimension, and vascular tortuosity are normalized to form multidimensional feature parameters. The trained classification model is then used to predict the probability of gestational hypertension and output the prediction results and the quantitative results of vascular morphology indicators. The classification model is SVM, logistic regression, random forest, or KNN algorithm.