Sclera image recognition and analysis method based on multi-modal features

Abstract

This invention discloses a scleral image recognition and analysis method based on multimodal features, belonging to the field of medical image processing technology. The method includes acquiring five-view original eye images of the subject, and using a trained deep neural network to filter qualified images; locating key points such as the inner and outer canthi and performing size normalization and orientation correction; establishing a polar coordinate system with the pupil center as the origin, dividing the sclera into clock positions and aggregating them into four quadrants to achieve position segmentation; identifying microvessels through a target detection model, extracting morphological features and assigning position codes; matching the feature and code data to a structured scoring knowledge base, calculating the health score deduction value for each position, and finally generating a health assessment report containing a visual annotation map and quantitative scores. This method can accurately segment the scleral position, achieve the binding scoring of microvessel features and position codes, and improve the accuracy and standardization of scleral image analysis.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically a method for scleral image recognition and analysis based on multimodal features. Background Technology

[0002] Existing technologies related to scleral image recognition and analysis mostly adopt a single-viewpoint approach to acquire eye images. They rely on manual methods to determine the integrity of the eyeball, image clarity, and the exposed area of ​​the sclera. They use a conventional planar coordinate system to simply divide the sclera position and extract scleral microvascular features through basic image processing. Eye health assessments rely on human experience to make judgments, and a standardized position segmentation model and structured scoring mechanism have not been formed.

[0003] Existing technologies have a single acquisition perspective, making it impossible to fully acquire image information of the entire sclera. There is no unified quantitative standard for manual quality judgment, which easily leads to deviations in image screening results. Conventional location segmentation methods do not match the physiological distribution characteristics of the sclera, resulting in low location segmentation accuracy. Microvascular features cannot be accurately correlated with corresponding physiological locations. Health scores lack fixed rule constraints, and the assessment results are obviously subjective, making it difficult to achieve refined segmentation of scleral locations and standardized quantitative analysis of microvascular features.

[0004] This invention requires establishing a polar coordinate system with the geometric center of the pupil as the origin, dividing the sclera into multiple clock positions and aggregating them into four quadrants to complete the position segmentation. At the same time, it extracts the morphological features of scleral microvessels and assigns corresponding position codes. The features and codes are matched with a preset structured scoring knowledge base to calculate the health score deduction value corresponding to each position. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a scleral image recognition and analysis method based on multimodal features, including: Acquire raw eye images of the subject from five different perspectives: frontal, upward, downward, leftward, and rightward. Using a trained deep neural network, the acquired raw eye images are automatically judged for eye integrity, image sharpness and effective scleral exposure area, and multi-view eye images that meet the preset quality standards are selected. For the selected multi-view eye images, the inner canthus, outer canthus, pupil geometric center and scleral contour are identified and located. Based on the location information, the image is normalized in size and corrected in orientation to generate a standard format multi-view eye analysis image. Using the geometric center of the pupil as the origin, a polar coordinate system is established on the standard format multi-view eye analysis image. The sclera position is divided into multiple clock points and further aggregated into four quadrants, thereby completing the position segmentation of the sclera. The target detection model is called to identify microvessels in the scleral images at each segmented location, extract multiple morphological features of each identified blood vessel, and call the multimodal large model to identify the location information of the blood vessels to obtain the location code. The data containing vascular morphological features and location codes are matched with a pre-set structured scoring knowledge base to calculate the health score deduction value corresponding to each location. A comprehensive analysis of the deduction values ​​in the health score is performed to generate a structured health assessment report that includes visual annotation charts and quantitative scores.

[0006] Furthermore, the step of using a trained deep neural network to automatically determine the integrity of the eyeball, image sharpness, and effective scleral exposure area of ​​the acquired raw eye images, and selecting multi-view eye images that meet preset quality standards, specifically includes the following steps: The original eye images of the subject acquired from five different perspectives—frontal, upward, downward, left, and right—are input into a pre-trained eye integrity discrimination model. This model, based on a convolutional neural network, outputs the probability value of the integrity of the eye structure in the image. A first integrity threshold is set, and images with integrity probability values ​​lower than this threshold are removed to filter out image data with excessive eyelid occlusion or missing main eyeballs. The image after initial screening for eyeball integrity is input into the image sharpness evaluation module. The sharpness evaluation module uses the Laplacian variance algorithm to calculate the gradient magnitude distribution of the image and combines it with the sharpness confidence output by the trained image quality classification network to comprehensively determine the image sharpness. A second sharpness threshold is set to remove blurry images that do not meet the sharpness requirements and retain image data with identifiable high-frequency details. For images that pass the clarity screening, a semantic segmentation network is used to accurately segment the scleral position mask, and the ratio of the mask pixel area to the total pixel area of ​​the entire eye image is calculated to obtain the effective scleral exposure rate. At the same time, the presence of obvious reflective interference positions in the image is detected, and if they exist, the area ratio of the reflective position to the scleral position is calculated. Set a minimum effective scleral exposure rate threshold and a maximum reflective interference ratio threshold, and remove images with an effective scleral exposure rate lower than the minimum threshold or a reflective interference ratio higher than the maximum threshold to exclude invalid images caused by shooting angle deviation or light source reflection. For each viewpoint, the single image with the highest comprehensive quality score is selected from the remaining images that meet all the above discrimination conditions as the representative image of that viewpoint. The comprehensive quality score is calculated by weighting the integrity probability value, the sharpness confidence value, and the effective scleral exposure rate, and finally forming a set of five high-quality, multi-view eye analysis images with consistent quality standards.

[0007] Furthermore, the step of establishing a polar coordinate system on the standard format multi-view eye analysis image with the pupil geometric center as the origin, and dividing the sclera position into multiple clock positions, specifically includes: In the standard format multi-view eye analysis image, the geometric center of the pupil position is calculated, and the geometric center of the pupil position is defined as the origin of the polar coordinate system; Centered on the origin, starting from the horizontal nasotemporal axis, the sclera is divided into a sector position every 30 degrees, and the sclera position is divided into twelve clock sector positions corresponding to the one to twelve o'clock positions on a clock. Calculate the range of pixel coordinates for each clock sector position in the standard format multi-view eye analysis image; Based on the actual position of the eyelid in the standard format multi-view eye analysis image, determine the observable portion actually exposed for each clock sector position, and mark the observable and unobservable positions. The twelve clock positions are further grouped into four analytical quadrants—upper left, upper right, lower left, and lower right—based on their orientation.

[0008] Furthermore, the target detection model is invoked to identify microvessels in the segmented scleral images at each location, extracting multiple morphological features of each identified vessel, specifically including: The segmented scleral image patches in each quadrant are then input into the optimized YOLO object detection model. The optimized YOLO object detection model identifies all microvascular structures in the image with a width greater than or equal to 0.05 mm, and outputs the bounding box coordinates and classification confidence of each microvascular. For each identified microvessel, five morphological features are extracted from its bounding box image location: spatial orientation, vessel diameter, color value, degree of vessel tortuosity, and whether there are spots at the vessel end. The extracted spatial orientation features are classified and assigned values ​​based on whether they point to the center of the pupil, whether they are broken, whether they point to other specific directions, or whether they are disordered. The extracted blood vessel diameter, color value, tortuosity, and terminal spot features are quantified and graded according to a preset threshold range.

[0009] Furthermore, the data containing vascular morphological features and location codes is matched against a pre-defined structured scoring knowledge base to calculate the health score deduction value corresponding to each location, specifically including: A structured scoring knowledge base containing multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights is pre-constructed; The spatial orientation level, vessel diameter level, color value level, tortuosity level, and terminal spot level of the vascular morphology features are compared item by item with the feature entries in the structured scoring knowledge base. When all feature dimensions match the rule entries in the knowledge base, the weighted deduction is calculated for the currently identified blood vessel based on the preset weight coefficients in the rule entries. The location code of the blood vessel is compared item by item with the feature entries in the structured scoring knowledge base. When an entry in the knowledge base is matched, a weighted deduction is calculated for the currently identified blood vessel location (corresponding to different organs) according to the preset weight coefficient in the rule entry.

[0010] Traverse all identified blood vessels within a specific location, accumulate the weighted deduction sum of all blood vessels within the specific location, and use it as the health score deduction value corresponding to the specific location; The health score deduction value corresponding to each location is bound to its location code to form a location-deduction value pair.

[0011] Furthermore, a comprehensive analysis is performed on the deduction values ​​of the health score to generate a structured health assessment report that includes visual annotation charts and quantitative scores, including: The system receives the calculated health score deduction values ​​for each location, combines them with the examinee's basic physiological parameters, and calls upon a personalized correction model to correct the scores. Based on the corrected location scores, the location scores are converted into corresponding human organ health evaluation values ​​through the organ mapping matrix; By combining the health evaluation values ​​of all organs, the overall health score of the nine major systems of the human body is calculated. Integrate organ health assessment values, systemic comprehensive health scores, and basic physiological parameters to generate a structured health assessment report that includes visual annotation charts and quantitative scores; The calculated health score deduction values ​​for each location are then combined with the examinee's basic physiological parameters to apply a personalized correction model to correct the scores. Specifically, this includes: The body mass index is calculated by obtaining the age, gender, height, and weight of the person being tested. The age, gender, and body mass index are used as inputs to feed into the personalized correction model. The personalized correction model incorporates baseline parameters based on population statistics, adjusts the age-related decay factor according to the input age value, selects the gender-specific coefficient based on gender information, and compensates for the intraocular pressure impact factor based on body mass index. The personalized correction model outputs age correction coefficients, gender correction coefficients, and body mass index correction coefficients for the test subject. The health score deduction for each location is multiplied by the age correction coefficient, gender correction coefficient, and body mass index correction coefficient to obtain the final corrected health score for each location.

[0012] Furthermore, based on the corrected location scores, the process of converting the location scores into corresponding human organ health evaluation values ​​through an organ mapping matrix specifically includes: Load a preset organ mapping matrix, which defines the correspondence between each quadrant of the sclera position and the clock position, and the specific organs in the nine major systems of the human body: digestive system, nervous system, respiratory system, cardiovascular system, musculoskeletal system, endocrine system, urinary system, male reproductive system, and female reproductive system. Based on the location code in the location-deduction value pair, query one or more target human organs associated with it in the organ mapping matrix; The final location health deduction value is allocated to each associated target human organ according to a preset mapping weight. For each target human organ, the total deduction value assigned to the target human organ from all associated locations is summed. The final health evaluation value of the target human organ is obtained by subtracting the sum of the deduction values ​​from the preset basic health score of the target human organ.

[0013] Furthermore, the comprehensive health assessment values ​​of all organs are used to calculate the overall health score of the nine major systems of the human body, specifically including: Based on the classification of human anatomy and physiology, all human organs are divided into nine major body systems; For each body system, obtain the final health assessment value of all constituent organs within that body system; Calculate the arithmetic mean of the final health evaluation values ​​of all organs under the body system, and use it as the initial comprehensive score of the body system; Based on the differences in the importance of different body systems, a preset system weight coefficient is assigned to each system. The initial comprehensive score of each body system is multiplied by its corresponding system weight coefficient, and then normalized to obtain the final comprehensive health score of the body system.

[0014] Furthermore, the integrated organ health assessment values, system comprehensive health scores, and basic physiological parameters generate a structured health assessment report containing visualized annotations and quantitative scores, specifically including: Create a structured health assessment report framework, which includes a report header, core conclusions area, test details area, organ health score area, body system assessment area, health recommendations area, and report footer; In the core conclusions section, enter the overall health score and health status level calculated based on the final comprehensive health score of the body system; Embed a multi-view standard image of eye analysis with vascular feature annotation in the detection details area; The organ health scoring area lists all assessed organs and their corresponding final health assessment values ​​and risk levels in tabular form. The final comprehensive health scores and comparative analysis of the nine major systems of the human body are displayed in chart form in the body system assessment area. In the health advice area, based on the organs and systems whose risk levels exceed the threshold, the corresponding standardized health advice text is retrieved from the preset knowledge base; Enter the test subject's basic physiological parameters, test date, and unique report number into the report header, and attach a standard disclaimer at the bottom of the report to complete the report generation.

[0015] Furthermore, the pre-constructed structured scoring knowledge base, which includes multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights, specifically includes: Collect and organize Zhuang medicine eye diagnosis theory and clinical observation data, and classify the five dimensions of blood vessel spatial orientation, blood vessel diameter, color, degree of tortuosity, and terminal spots into three or four abnormal levels respectively; Each anomaly level feature is assigned a quantified feature code, and a base deduction value is set for features of different levels. Establish the correlation between feature coding, basic deduction values, and specific quadrants and clock positions of the sclera; Establish mapping relationships between specific quadrants of the sclera, clock positions, and human organs; Feature encoding, basic deduction values, quadrant point associations, organ mapping associations, and disease weight coefficients for specific organs are integrated into structured JSON format knowledge entries and stored as a structured scoring knowledge base file.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Using the geometric center of the pupil as the origin, a polar coordinate system is established on a standard format multi-view eye analysis image. The sclera is divided according to the clock position, and the divided points are further aggregated into four quadrants to complete the sclera position segmentation. This method can adapt the sclera position segmentation method to the physiological structure of the eye, accurately determine the spatial range of different scleral zones, clarify the boundary information of each segmentation position, and avoid the problems of vague boundary definition and mismatch between the zone and the physiological structure that exist in conventional position segmentation methods. This makes the spatial assignment of each scleral position clearly distinguishable, realizes the fine and standardized segmentation of sclera position, and ensures that scleral images from different perspectives can form a unified and standardized position segmentation result, guaranteeing the consistency and accuracy of position segmentation.

[0017] By using a target detection model to identify microvessels in the segmented scleral images at various locations, multiple morphological features corresponding to a single microvessel are extracted. Each extracted vascular feature is matched with a corresponding location code, ensuring a unique correspondence between the microvessel feature and its corresponding scleral location. The morphological feature data carrying the location code is then matched with the execution rules of a pre-defined structured scoring knowledge base to directly calculate the health score deduction value for each location. This avoids the problems of inconsistent standards and strong subjectivity in results caused by manual assessment, ensuring that the score deduction calculation process follows fixed structured rules. The score deduction value for each location can directly reflect the actual characteristic state of the microvessels within the corresponding location, achieving full standardization of microvessel feature extraction, location association, and score calculation. This provides a unified execution logic for the health score generation process, ensuring the objectivity and stability of the assessment results. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the scleral image recognition and analysis method based on multimodal features described in this invention. Figure 2 A time-series graph used to call a target detection model to identify microvessels and extract morphological features; Figure 3 This is a scleral clock position-organ mapping weight coefficient diagram; Figure 4 A line graph showing the correction and comparison of the scleral four-quadrant health score; Figure 5 A weighted heatmap mapping the sclera quadrant to the nine major systems of the human body. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention provides a method for scleral image recognition and analysis based on multimodal features, the specific method including: The system acquires raw eye images of the subject from five different perspectives: frontal, upward, downward, leftward, and rightward. A trained deep neural network is used to automatically assess the integrity of the eyeball, image clarity, and effective scleral exposure area of ​​the acquired raw eye images, selecting multi-view eye images that meet preset quality standards. For the selected multi-view eye images, the inner and outer canthi, the geometric center of the pupil, and the scleral contour are identified and located. Based on the location information, the images are normalized in size and corrected in orientation to generate a standard-format multi-view eye analysis image. A polar coordinate system is established on the standard-format multi-view eye analysis image with the geometric center of the pupil as the origin, dividing the scleral position into multiple clock positions, which are further aggregated into four quadrants, thus completing the scleral position segmentation. A target detection model is invoked to identify microvessels in the segmented scleral images, extracting multiple morphological features of each identified blood vessel. A multimodal large model is then used to identify the location information of the blood vessels to obtain positional codes. Data containing vascular morphological features and location codes are matched against a pre-defined structured scoring knowledge base using rules to calculate the corresponding health score deduction value for each location. A comprehensive analysis of these health score deduction values ​​generates a structured health assessment report that includes a visually labeled map and a quantitative score.

[0021] In one embodiment of the present invention, the original eye images of the subject acquired from five different perspectives (frontal, upward, downward, left, and right) are input into a pre-trained eye integrity discrimination model. This model, based on a convolutional neural network, outputs the probability value of the integrity of the eye structure in the image. A first integrity threshold is set, and images with integrity probability values ​​below this threshold are removed to filter out image data with excessive eyelid occlusion or missing main eye structures. The images after initial screening for eye integrity are input into an image sharpness evaluation module. This module uses the Laplacian variance algorithm to calculate the gradient magnitude distribution of the image and combines it with the sharpness confidence score output by a trained image quality classification network to comprehensively determine the image sharpness. A second sharpness threshold is set, and blurry images that do not meet the sharpness requirements are removed, retaining image data with identifiable high-frequency details. For images that pass the sharpness screening, a semantic segmentation network is used to accurately segment the scleral mask, and the ratio of the mask pixel area to the total pixel area of ​​the entire eye image is calculated to obtain the effective scleral exposure rate. Simultaneously, the system detects whether there are obvious reflective interference locations in the image. If so, it calculates the area ratio of the reflective location to the sclera. A minimum effective scleral exposure rate threshold and a maximum reflective interference ratio threshold are set. Images with an effective scleral exposure rate lower than the minimum threshold or a reflective interference ratio higher than the maximum threshold are removed to eliminate invalid images caused by shooting angle deviations or light source reflections. For each viewpoint, the single image with the highest overall quality score is selected from the remaining images that meet all the above criteria as the representative image for that viewpoint. The overall quality score is calculated by weighting the integrity probability value, sharpness confidence value, and effective scleral exposure rate, ultimately forming a set of five high-quality, multi-view eye analysis images with consistent quality standards.

[0022] In practice, the automated discrimination and screening process based on multi-view eye images is executed as follows: After acquiring the original eye images of the subject from five different perspectives (frontal, upward, downward, left, and right), the processing system inputs each original eye image into a pre-trained eye integrity discrimination model. The eye integrity discrimination model is built on a convolutional neural network architecture, and its output is a scalar value between 0 and 1. This scalar value represents the probability that the eye structure in the image is intact. In practice, the first integrity threshold is set to 0.85. Any image with an integrity probability value lower than 0.85 will be automatically rejected. This step effectively filters out image data that is missing or severely obscured by the main body of the eye due to excessive eyelid closure, blinking, or shooting errors. After the initial screening for eyeball integrity is completed, the remaining images are sent to the image sharpness evaluation module. The image sharpness evaluation module uses the Laplacian variance algorithm to calculate the gradient magnitude variance of the entire image. At the same time, a parallel-running, trained image quality classification network outputs a sharpness confidence score for the same image. The sharpness evaluation module integrates the gradient magnitude variance and the sharpness confidence score to obtain a comprehensive sharpness evaluation value. In specific implementation, a second sharpness threshold is set as a preset comprehensive evaluation value threshold. Any image with a comprehensive sharpness evaluation value lower than this threshold is identified as a blurry image and discarded, thereby retaining image data with sufficiently clear high-frequency details that are convenient for subsequent microvascular identification.

[0023] For images that pass the clarity screening, the system calls a trained semantic segmentation network. The semantic segmentation network generates a binary scleral location mask for the input image, where pixels in the scleral location are marked as 1 and pixels in non-scleral locations are marked as 0. The total number of pixels with a value of 1 in the scleral location mask is calculated and divided by the total number of pixels in the entire eye image to obtain the effective scleral exposure rate. At the same time, the image processing algorithm detects whether there are high-brightness pixel clusters within the scleral location and identifies them as reflective interference locations. The proportion of the pixel area of ​​the reflective interference location to the total pixel area of ​​the scleral location is calculated. In specific implementation, the minimum effective scleral exposure rate threshold is set to 15%, and the maximum reflective interference ratio threshold is set to 10%. Any image with an effective scleral exposure rate lower than 15% or a reflective interference ratio higher than 10% will be excluded. This rule aims to exclude images where the sclera is exposed too little due to improper shooting angle, or where ambient light directly shines on the sclera, forming a bright spot that affects the observation of blood vessels.

[0024] In some embodiments, for each specified viewpoint, after passing all the above quality checks, multiple candidate images may still remain. The system needs to select the best representative image from these. In a specific implementation, a comprehensive quality score is calculated for each candidate image. The formula for calculating the comprehensive quality score is as follows: in: Represents the overall quality score. This represents the integrity probability value output by the eyeball integrity discrimination model. The sharpness confidence score represents the output of the image quality classification network. This represents the calculated effective scleral exposure rate. , , These are preset weighting coefficients that satisfy... In practical implementation, the weighting coefficient , , The values ​​can be set to 0.4, 0.4, and 0.2 respectively, and the system calculates the overall quality score of all candidate images at each viewpoint. , and select The image with the highest value is used as the final representative image for that viewpoint. Through this process, the system outputs a set of high-quality eye analysis images for each eye, including five viewpoints: frontal, upward, downward, leftward, and rightward. This set of images has a consistent quality standard.

[0025] In one embodiment of the present invention, in the standard format multi-view eye analysis image, the geometric center of the pupil position is calculated, and the geometric center of the pupil position is defined as the origin of the polar coordinate system. Centered on the origin, starting from the horizontal nasotemporal axis, the image is divided into 30-degree sectors, resulting in a total of twelve clock-shaped sectors corresponding to the positions of one to twelve o'clock on a clock face. The pixel coordinate range of each clock-shaped sector in the standard format multi-view eye analysis image is calculated. Based on the actual position of the eyelid in the standard format multi-view eye analysis image, the observable portion of each clock-shaped sector is determined, and observable and unobservable positions are marked. The twelve clock-shaped positions are further aggregated into four analysis quadrants: upper left, upper right, lower left, and lower right, according to their orientation. The segmented scleral image blocks of each quadrant are input into an optimized and trained YOLOv8 object detection model. The optimized YOLOv8 object detection model identifies all microvascular structures in an image with a width greater than or equal to 0.05 mm, outputting the bounding box coordinates and classification confidence score for each microvessel. For each identified microvessel's bounding box image location, five morphological features are extracted: spatial orientation, vessel diameter, color value, vessel curvature, and the presence of spots at the vessel's terminal. The extracted spatial orientation features are classified and assigned values ​​based on whether they point towards the pupil center, are interrupted, point in another specific direction, or are disordered. The extracted vessel diameter, color value, curvature, and terminal spot features are quantized and graded according to preset threshold ranges.

[0026] In practice, the specific process of scleral location segmentation and vascular feature extraction is performed as follows: The processing system operates on the standard format multi-view eye analysis image that has undergone size normalization and orientation correction. In the standard format multi-view eye analysis image, the geometric center of the pupil position is calculated using an image processing algorithm. The algorithm calculates the arithmetic mean of the horizontal and vertical coordinates of all pixels on the pupil outline to obtain the coordinates of the pupil geometric center, and defines these coordinates as the origin of the polar coordinate system. With the pupil geometric center as the origin, a horizontal ray pointing to the right is defined as the polar axis, corresponding to the 3 o'clock position on a clock face. Starting from this polar axis, a sector position is divided every 30 degrees counterclockwise, thus uniformly dividing the scleral location into 12 clock sector positions corresponding to the 12 o'clock, 11 o'clock, 10 o'clock, and up to the 1 o'clock position on a clock face. The pixel coordinate range of each clock sector position in the standard format multi-view eye analysis image is calculated. Based on the conversion relationship between polar coordinates and rectangular coordinates, the system determines the boundary of each 30-degree sector in the image pixel plane. Based on the segmented eyelid contour position information, the system determines the observable portion actually exposed for each clock-shaped sector. Sectors obscured by the upper or lower eyelid are marked as unobservable, while only the sclera portion not obscured by the eyelids is marked as observable for that clock position. The 12 clock positions are aggregated according to their orientation: 10, 11, and 12 o'clock are aggregated into the upper left quadrant; 1, 2, and 3 o'clock into the upper right quadrant; 4, 5, and 6 o'clock into the lower right quadrant; and 7, 8, and 9 o'clock into the lower left quadrant.

[0027] In some embodiments, after location segmentation is completed, the system calls an optimized YOLOv8 object detection model to perform blood vessel identification, see [reference]. Figure 2The scleral image patches obtained from the segmentation of the upper left, upper right, lower left, and lower right quadrants are input into the YOLOv8 object detection model. The optimized YOLOv8 model is configured to identify all microvascular structures in the image with a width greater than or equal to 0.05 mm. The YOLOv8 model outputs the bounding box coordinates and a classification confidence score for each identified microvessel. For the bounding box image location of each microvessel output by the YOLOv8 model, the system extracts five morphological features: spatial orientation, vessel diameter, color value, vessel tortuosity, and the presence of spots at the vessel tip. The extracted spatial orientation features are classified and assigned values ​​according to preset rules, defined as follows: if the main extension direction of the blood vessel points towards the geometric center of the pupil, it is assigned a value of A1; if the blood vessel shows discontinuous breaks, it is assigned a value of A2; if the blood vessel points towards a specific direction such as the temporal side, nasal side, upper side, or lower side, it is assigned a value of A3; if the blood vessel orientation is chaotic and without obvious pattern, it is assigned a value of A4. The extracted blood vessel diameter, color value, blood vessel curvature, and terminal spot features are quantified and graded according to preset threshold ranges. The blood vessel diameter is graded based on the physical measurement value converted from its average pixel width; the color value is graded based on the specific channel value of the blood vessel position in the RGB or HSV color space; the blood vessel curvature is graded by calculating the curvature or curvature index of the curve; and the presence of spots at the blood vessel terminal is determined by binarizing the pixel texture features at the blood vessel terminal position.

[0028] Optionally, when extracting vessel diameter features, the system samples at equal intervals along the centerline of the identified vessels and calculates the vessel width perpendicular to the centerline at each sampling point. The average width of all sampling points is then taken as the diameter feature value of the vessel. When extracting vessel color features, the system calculates the average intensity difference between the red and green channels of the vessel pixels within the bounding box and uses this average as the color feature. When extracting vessel tortuosity features, the system calculates the tortuosity index of the vessel centerline. The formula for calculating the tortuosity index is: in: Represents the curvature index. The actual length of the curve representing the centerline of the blood vessel. The tortuosity index represents the straight-line distance between the origin and end of a blood vessel. The higher the value, the greater the tortuosity of the blood vessel. This is understandable, based on the tortuosity index... The system maps the value to a preset bending level, for example... To be straight, Slightly bent. For significant curvature. When extracting speckle features at the end of the blood vessel, the system calculates the texture contrast and uniformity of the local image within a small neighborhood of the end of the blood vessel. If a location with high contrast and an approximately circular shape is detected, speckle features are determined to exist.

[0029] In some embodiments, after assigning a level to each morphological feature, the system generates a feature vector for each identified blood vessel. The feature vector includes a spatial orientation classification value, vessel diameter level, color value level, vessel tortuosity level, and an indicator of the presence or absence of terminal spots. Simultaneously, the system assigns a location code to the vessel based on its position. The location code is composed of a viewpoint code, a quadrant code, and the corresponding clock position number. It can be understood that all identified blood vessels and their corresponding multidimensional morphological features and location codes are organized into a structured data list for use by subsequent rule matching and scoring calculation modules.

[0030] In one embodiment of the present invention, a structured scoring knowledge base containing multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights is pre-constructed. Zhuang medicine eye diagnosis theory and clinical observation data are collected and organized, and the five dimensions of vascular spatial orientation, vascular diameter, color, curvature, and terminal spots are divided into three or four abnormality levels. A quantified feature code is assigned to each abnormality level, and a basic deduction value is set for different levels of features. The correlation between feature codes, basic deduction values, and specific quadrants and clock positions of the sclera is established. A mapping correlation between specific quadrants and clock positions of the sclera and human organs is established. The feature codes, basic deduction values, quadrant position correlations, organ mapping correlations, and disease weight coefficients for specific organs are integrated into structured JSON format knowledge entries and stored as a structured scoring knowledge base file. The spatial orientation level, vascular diameter level, color value level, curvature level, and terminal spot level of the vascular morphology features are compared item by item with the feature entries in the structured scoring knowledge base. When all feature dimensions match rule entries in the knowledge base, a weighted deduction is calculated for the currently identified blood vessel based on the preset weight coefficients in the rule entries. All identified blood vessels within a specific location are traversed, and the total weighted deduction for all blood vessels within that specific location is accumulated as the health score deduction value corresponding to that specific location. The health score deduction value corresponding to each location is then bound to its location code, forming a location-deduction value pair.

[0031] In practice, the construction of the structured scoring knowledge base and the calculation of health score deductions are performed as follows: A structured scoring knowledge base containing multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights is pre-constructed. Zhuang medicine eye diagnosis theory and clinical observation data are collected and organized. The spatial orientation dimension of blood vessels is divided into four abnormality levels: extending towards the pupil, interrupted, extending to other areas, and disordered. The size dimension is divided into three abnormality levels: large at the root, moderately large, and small. The color dimension is divided into three abnormality levels: deep red / purple, bright red, and light red. The curvature dimension is divided into three abnormality levels: spiral, serpentine, relatively straight, or irregular. The terminal spot dimension is divided into three levels: plaque / large spot, small spot, and no spot. Each abnormality level is assigned a quantified feature code and a basic deduction value. For example, the spatial orientation feature of extending towards the pupil is coded as 3, with a basic deduction of 3 points. Simultaneously, based on Zhuang medicine eye diagnosis theory, an independent mapping relationship is established between vascular location codes and human organs. For example, the 10 o'clock position of the left eye is mapped to the liver. The aforementioned vascular morphology feature scoring rules, vascular location coding mapping rules, and location weight coefficients for different mapped organs are integrated into structured JSON format knowledge entries and stored as a local file to form a structured scoring knowledge base file. The core content of this knowledge base is illustrated in Table 1.

[0032] Table 1: Examples of the mapping between vascular morphological feature scoring rules and vascular location coding (location-organ). In practice, the system performs rule matching and scoring calculation, receiving the multidimensional morphological feature vector and corresponding location code for each identified blood vessel. The morphological feature vector includes spatial orientation level, vessel diameter level, color value level, tortuosity level, and terminal spot level. The spatial orientation level, vessel diameter level, color value level, tortuosity level, and terminal spot level from the vessel morphological feature vector are compared item by item with feature entries in the structured scoring knowledge base. The matching process involves finding entries in the knowledge base whose feature codes completely match the current blood vessel's feature levels. When all feature dimensions match rule entries in the knowledge base, a weighted scoring calculation is performed on the currently identified blood vessel based on the preset weight coefficients in the rule entries. The weighted scoring calculation formula is as follows: in: This represents the weighted deduction value for a single blood vessel. This represents the base deduction value defined in the matched feature entries. This represents the quadrant position weight coefficient retrieved from the knowledge base based on the quadrant and clock position of the blood vessel. This represents the disease weight coefficient retrieved from the knowledge base based on the mapped organ. It iterates through all identified blood vessels within a specific location, accumulating the weighted deductions for all vessels at that location, which serves as the deduction value for the health score corresponding to that location. This accumulation process can be understood as being performed for each independent analysis location; for example, for the location "left eye_frontal vision_upper left quadrant," the weighted deduction value for all identified blood vessels within that location is calculated. Add them together to get the total deduction value for that position. Deduct the corresponding health score value for each location. This data is bound to its location code to form a "location-deduction value pair" data structure. In some embodiments, the location-deduction value pairs are stored in the form of a list, where each item contains a location code string and a corresponding deduction value floating-point number. This list serves as input data for the subsequent personalized scoring correction module.

[0033] Optional, quadrant point weighting coefficients in weighted deduction calculation and disease weight coefficient The value ranges from 0.5 to 2.0. The magnitude of the coefficient reflects the importance of the feature at a specific location or for a specific organ. The weight coefficients are pre-set by domain experts based on clinical experience and stored in the structured scoring knowledge base. It can be understood that the structured scoring knowledge base is a scalable, independent module. Updates to its content, such as adding feature entries, adjusting basic deduction values, or modifying weight coefficients, do not affect the aforementioned image processing and feature extraction process; only the knowledge base file needs to be replaced to update the scoring rules. In some embodiments, when a blood vessel matches multiple feature entries simultaneously, such as a dark red blood vessel with significant curvature, the system will calculate the weighted deductions corresponding to the color feature and the curvature feature separately, and add them together as the total deduction value for this blood vessel, which is then included in the location deduction accumulation.

[0034] See Figure 3This is a scleral clock position-organ mapping weight coefficient diagram. Its core function is to quantify the importance of the correlation between different scleral locations and corresponding human organs. It transforms the traditional Zhuang medicine theory of "the white of the eye corresponding to internal organs" into quantifiable weight coefficients, allowing for an interpretable mapping relationship between traditional Chinese medicine theory and modern AI algorithms. This weight is the core parameter of the weighted deduction formula, directly determining the degree of influence of vascular features at different scleral locations on organ health scores, and is a key visual basis for the algorithm's interpretability. By determining the weight, the scleral locations that need to be prioritized in health assessments are identified, providing priority guidance for clinical testing. The weight coefficients are set by domain experts based on Zhuang medicine eye diagnosis theory and clinical data, ranging from 0.5 to 2.0, reflecting the importance of organs corresponding to different scleral locations. Vascular abnormalities at high-weight locations have a greater deduction impact on organ health scores and are the core monitoring locations for health assessment. Vascular abnormalities at low-weight locations have a smaller impact on the overall score and are used for auxiliary assessment.

[0035] In one embodiment of the present invention, the calculated health score deduction values ​​for each location are received, and combined with the examinee's basic physiological parameters, a personalized correction model is invoked to correct the scores. The examinee's age, gender, height, and weight are obtained, and the body mass index (BMI) is calculated. The age, gender, and BMI are input into the personalized correction model. The personalized correction model incorporates baseline parameters based on population statistics, adjusts the age-related decay factor based on the input age, selects a gender-specific coefficient based on the gender information, and compensates for intraocular pressure influence factors based on the BMI. The personalized correction model outputs age correction coefficients, gender correction coefficients, and BMI correction coefficients for the examinee. The health score deduction values ​​for each location are multiplied by the age correction coefficients, gender correction coefficients, and BMI correction coefficients, respectively, to obtain the corrected final location health deduction value. A preset organ mapping matrix is ​​loaded, which defines the correspondence between each quadrant and clock position of the sclera and specific organs in nine major systems of the human body: digestive, nervous, respiratory, cardiovascular, musculoskeletal, endocrine, urinary, male, and female reproductive systems. Based on the position code in the position-deduction value pair, one or more target human organs associated with the location are queried in the organ mapping matrix. The final position health deduction value is allocated to each associated target human organ according to a preset mapping weight. For each target human organ, the sum of the deduction values ​​allocated to it from all associated positions is summarized. The sum of the deduction values ​​is subtracted from the preset basic health score of the target human organ to obtain its final health evaluation value. Based on human anatomy and physiology, all human organs are classified into the nine major body systems. For each body system, the final health evaluation value of all constituent organs within that system is obtained. The arithmetic mean of the final health evaluation values ​​of all organs within that body system is calculated as the initial comprehensive score for that body system. Based on the differences in the importance of different body systems, a preset system weight coefficient is assigned to each system. The initial comprehensive score of each body system is multiplied by its corresponding system weight coefficient, and then normalized to obtain the final comprehensive health score of the body system.

[0036] In practice, the personalized correction, organ mapping, and system comprehensive score calculation for health scores are performed as follows: The calculated health score deduction values ​​for each location are received, and combined with the examinee's basic physiological parameters, a personalized correction model is invoked to correct the scores. The examinee's age, gender, height, and weight are obtained, and the Body Mass Index (BMI) is calculated by dividing the weight by the square of the height. The age, gender, and BMI are input into the personalized correction model. The model incorporates baseline parameters based on large-scale population statistics, adjusting the age-related attenuation factor based on the input age, selecting the gender-specific coefficient based on the gender information, and compensating for intraocular pressure influence factors based on the BMI. The personalized correction model outputs age correction coefficients, gender correction coefficients, and BMI correction coefficients for the examinee. The health score deduction values ​​for each location are multiplied by these coefficients to obtain the final corrected health score deduction value for each location. An example of the relationship between personalized correction coefficients and physiological parameters is shown in Table 2.

[0037] Table 2: Examples of the Relationship between Personalized Correction Coefficients and Physiological Parameters The formula for calculating the final health deduction value after correction is as follows: in: This represents the final, corrected positional health deduction value. This represents the uncorrected location health score deduction. Represents the age correction factor. Represents the gender correction factor. This represents the body mass index correction factor. A preset organ mapping matrix is ​​loaded. This matrix is ​​a data structure that defines the correspondence between each quadrant of the sclera location and its clock position and specific organs in the nine major human systems: digestive, nervous, respiratory, cardiovascular, musculoskeletal, endocrine, urinary, male reproductive, and female reproductive systems. Based on the location code in the location-deduction value pair, one or more target human organs associated with that location are queried in the organ mapping matrix, and the final location health deduction value is calculated. Based on preset mapping weights, the values ​​are assigned to each associated target human organ. For each target human organ, the total deduction values ​​assigned to all associated locations are summed. The total deduction values ​​are subtracted from the preset basic health score of the target human organ to obtain the final health evaluation value of the target human organ. The preset basic health score of the target human organ is 100 points by default.

[0038] In some embodiments, based on human anatomy and physiology, all human organs are divided into nine major body systems. For each body system, the final health evaluation value of all constituent organs within that system is obtained. The arithmetic mean of the final health evaluation values ​​of all organs within a body system is calculated as the initial comprehensive score for that body system. A preset system weight coefficient is assigned to each system based on their relative importance. The initial comprehensive score of each body system is multiplied by its corresponding system weight coefficient, and then normalized to obtain the final comprehensive health score for that body system. An example of calculating the system weight coefficient and normalization is as follows: Assuming the initial comprehensive score of the cardiovascular system is 85 points, its system weight coefficient is preset to 1.2; the initial comprehensive score of the musculoskeletal system is 90 points, its system weight coefficient is preset to 1.0; and normalization transforms all weighted scores to a range of 0-100, the final comprehensive health score for the cardiovascular system might be 88 points, and the final comprehensive health score for the musculoskeletal system might be 87 points.

[0039] Optionally, the system weight coefficients are set with reference to clinical medical consensus; for example, the weight coefficients for the cardiovascular and nervous systems are typically higher than those for the musculoskeletal and urinary systems. It is understood that the organ mapping matrix, mapping weights, and system weight coefficients are all stored in configuration files, allowing for adjustments and updates based on the latest medical research findings without modifying the core calculation program code. In some embodiments, different versions of the organ mapping matrix are loaded for male and female test subjects. The organ mapping matrix for female test subjects includes the organ mapping relationships of the female reproductive system, while the organ mapping matrix for male test subjects includes the organ mapping relationships of the male reproductive system. It is understood that both the final health evaluation value of the organ and the final comprehensive health score of the system are presented as percentages; a higher score indicates a better health status assessment of the organ or system under the current evaluation model.

[0040] See Figure 4 This is a line graph showing the correction and comparison of the scleral four-quadrant health score. The original deduction value is the location deduction value directly calculated from the morphological characteristics of the scleral microvessels. The corrected deduction value is the final deduction value after being adjusted by a personalized correction model, taking into account the examinee's age, gender, BMI, and other physiological parameters. It clearly demonstrates the effect of personalized correction on the original score, proving that the model can adjust the scleral location deduction value according to individual physiological characteristics, overcoming the limitations of a uniform scoring standard. The upper right quadrant shows the highest deduction value, indicating the most significant vascular abnormality at that location, which can further pinpoint the health risk of the corresponding organ. The higher the corrected deduction value, the higher the health risk of the corresponding organ, which can be used to guide personalized health assessment and intervention recommendations, aligning with the patent's goal of generating a structured health report.

[0041] In one embodiment of the present invention, a structured health assessment report framework is created, comprising a report header, a core conclusion area, a test details area, an organ health score area, a body system assessment area, a health advice area, and a report footer. The core conclusion area contains the overall health score and health status level calculated based on the final comprehensive health score of the body systems. The test details area embeds a multi-view standard image of the eye with vascular feature annotation. The organ health score area lists all assessed organs and their corresponding final health evaluation values ​​and risk levels in tabular form. The body system assessment area displays the final comprehensive health scores and comparative analysis of the nine major human systems in chart form. In the health advice area, standardized health advice text is retrieved from a preset knowledge base for organs and systems with risk levels exceeding a threshold. The tester's basic physiological parameters, test date, and unique report number are entered into the report header, and a standard disclaimer is attached to the report footer to complete the report generation.

[0042] In practice, the generation process of the structured health assessment report is carried out as follows: A structured health assessment report framework is created, using an electronic document template. The template pre-divides seven fixed positions: report header, core conclusion area, test details area, organ health score area, body system assessment area, health recommendations area, and report bottom. The core conclusion area is filled with the overall health score and health status level calculated based on the final comprehensive health score of the body systems. The formula for calculating the overall health score is: in: Represents the overall health score. Representing the The final comprehensive health score of an individual's bodily systems. Representing the The system weight coefficients of each body system are preset, and the summation covers all nine major body systems. The result is calculated as follows: After the value is calculated, the health status level is determined based on a preset threshold range, for example... "Excellent" For "good", The "Focus" setting is used to embed multi-view standard images of the eye, annotated with vascular features, in the detection details area. These images are generated from the aforementioned process and are five standard eye images (frontal, upward, downward, left, and right views) with different colors and markers indicating vascular morphology and their characteristic categories. The organ health scoring area lists all assessed organs and their corresponding final health evaluation values ​​and risk levels in tabular form. The table includes at least three columns: "Organ Name," "Health Evaluation Value," and "Risk Level." The risk level is automatically determined based on the preset range of the final health evaluation value; for example, a score above 80 is "low risk," between 60 and 80 is "medium risk," and below 60 is "high risk."

[0043] In some embodiments, the final comprehensive health scores and comparative analysis of the nine major human systems are displayed in a chart format in the body system assessment area. The chart includes a horizontal bar chart showing the final comprehensive health scores of the digestive system, nervous system, respiratory system, cardiovascular system, musculoskeletal system, endocrine system, urinary system, male reproductive system, and female reproductive system, respectively. An average score for each system, calculated based on historical population data, is attached next to the chart as a reference baseline. In the health advice area, the system retrieves standardized health advice text from a pre-defined knowledge base for organs and systems with risk levels exceeding a threshold. The retrieval logic involves iterating through the organ health scoring table and body system assessment results, identifying all items with risk levels of "medium risk" and "high risk," matching and extracting predefined text advice from the knowledge base based on the item name, and then categorizing these text advice by organ and system and filling them into the health advice area. The tester's basic physiological parameters, test date, and unique report number are entered into the report header. The report number is generated using the rule "test date + sequence number," and a standard disclaimer is attached to the bottom of the report, completing the report generation.

[0044] Optionally, the health assessment report framework supports output in PDF or HTML webpage format. PDF is used for printing and archiving, while HTML is used for interactive display on user terminals. Users can click on specific items in charts or tables on the webpage to view more detailed explanations. It is understood that the multi-view ocular analysis standard images embedded in the report, annotated with vascular features, have annotation information completely consistent with the data from previous feature extraction and rule matching stages, ensuring the correspondence between the report content and the underlying analysis data. In some embodiments, the core conclusion area, in addition to displaying the overall health score and level, is visually enhanced with prominent icons or color blocks and includes a brief summary description. Optionally, the standardized health advice text library in the health advice area allows administrators to remotely update and maintain it according to the latest medical guidelines or product positioning without modifying the report generation program itself. It is understood that the entire report generation process is automated, from receiving the corrected organ and system scores to finally generating a complete structured document, without manual intervention. After generation, the report is pushed to the user terminal or stored in the server database via a designated interface.

[0045] See Figure 5 This is a heatmap mapping the sclera quadrants to the nine major systems of the human body, its core function being to quantify the correlation strength between different locations of the sclera and various systems in the body. It transforms the empirical theory of "the sclera belonging to different organs" in Zhuang medicine's eye diagnosis into a quantifiable two-dimensional weight matrix, realizing the digitization and algorithmization of TCM theory. This weight matrix is ​​a core component of the organ mapping matrix, directly determining the allocation ratio of scleral position deductions to system scores, and is a key visual basis for the algorithm's interpretability. The heatmap visually locates high-weight associations, clearly identifying the quadrant-system combinations that need to be analyzed in health assessments, providing priority guidance for clinical testing. The weight coefficients, set by domain experts based on Zhuang medicine's eye diagnosis theory and clinical data, range from 0.1 to 0.9, reflecting the importance of the systems corresponding to different scleral quadrants.

[0046] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for scleral image recognition and analysis based on multimodal features, characterized in that, Includes the following steps: Acquire raw eye images of the subject from five different perspectives: frontal, upward, downward, leftward, and rightward. Using a trained deep neural network, the acquired raw eye images are automatically judged for eye integrity, image sharpness and effective scleral exposure area, and multi-view eye images that meet the preset quality standards are selected. For the selected multi-view eye images, the inner canthus, outer canthus, pupil geometric center and scleral contour are identified and located. Based on the location information, the image is normalized in size and corrected in orientation to generate a standard format multi-view eye analysis image. Using the geometric center of the pupil as the origin, a polar coordinate system is established on the standard format multi-view eye analysis image. The sclera position is divided into multiple clock points and further aggregated into four quadrants, thereby completing the position segmentation of the sclera. The target detection model is called to identify microvessels in the scleral images at each segmented location, extract multiple morphological features of each identified blood vessel, and call the multimodal large model to identify the location information of the blood vessels to obtain the location code. The data containing vascular morphological features and location codes are matched with a pre-set structured scoring knowledge base to calculate the health score deduction value corresponding to each location. A comprehensive analysis of the deduction values ​​in the health score is performed to generate a structured health assessment report that includes visual annotation charts and quantitative scores.

2. The scleral image recognition and analysis method based on multimodal features according to claim 1, characterized in that, The process utilizes a trained deep neural network to automatically assess the integrity of the eyeball, image sharpness, and effective scleral exposure area of ​​the acquired raw eye images, selecting multi-view eye images that meet preset quality standards. This specifically includes the following steps: The original eye images of the subject acquired from five different perspectives—frontal, upward, downward, left, and right—are input into a pre-trained eye integrity discrimination model. This model, based on a convolutional neural network, outputs the probability value of the integrity of the eye structure in the image. A first integrity threshold is set, and images with integrity probability values ​​lower than this threshold are removed to filter out image data with excessive eyelid occlusion or missing main eyeballs. The image after initial screening for eyeball integrity is input into the image sharpness evaluation module. The sharpness evaluation module uses the Laplacian variance algorithm to calculate the gradient magnitude distribution of the image and combines it with the sharpness confidence output by the trained image quality classification network to comprehensively determine the image sharpness. A second sharpness threshold is set to remove blurry images that do not meet the sharpness requirements and retain image data with identifiable high-frequency details. For images that pass the clarity screening, a semantic segmentation network is used to accurately segment the scleral position mask, and the ratio of the mask pixel area to the total pixel area of ​​the entire eye image is calculated to obtain the effective scleral exposure rate. At the same time, the presence of obvious reflective interference positions in the image is detected, and if they exist, the area ratio of the reflective position to the scleral position is calculated. Set a minimum effective scleral exposure rate threshold and a maximum reflective interference ratio threshold, and remove images with an effective scleral exposure rate lower than the minimum threshold or a reflective interference ratio higher than the maximum threshold to exclude invalid images caused by shooting angle deviation or light source reflection. For each viewpoint, the single image with the highest comprehensive quality score is selected from the remaining images that meet all the above discrimination conditions as the representative image of that viewpoint. The comprehensive quality score is calculated by weighting the integrity probability value, the sharpness confidence value, and the effective scleral exposure rate, and finally forming a set of five high-quality, multi-view eye analysis images with consistent quality standards.

3. The scleral image recognition and analysis method based on multimodal features according to claim 2, characterized in that, The process involves establishing a polar coordinate system on the standard format multi-view eye analysis image, with the geometric center of the pupil as the origin, and dividing the sclera position into multiple clock positions. Specifically, this includes: In the standard format multi-view eye analysis image, the geometric center of the pupil position is calculated, and the geometric center of the pupil position is defined as the origin of the polar coordinate system; Centered on the origin, starting from the horizontal nasotemporal axis, the sclera is divided into a sector position every 30 degrees, and the sclera position is divided into twelve clock sector positions corresponding to the one to twelve o'clock positions on a clock. Calculate the range of pixel coordinates for each clock sector position in the standard format multi-view eye analysis image; Based on the actual position of the eyelid in the standard format multi-view eye analysis image, determine the observable portion actually exposed for each clock sector position, and mark the observable and unobservable positions. The twelve clock positions are further grouped into four analytical quadrants—upper left, upper right, lower left, and lower right—based on their orientation.

4. The scleral image recognition and analysis method based on multimodal features according to claim 3, characterized in that, The target detection model is invoked to identify microvessels in the segmented scleral images at each location, and to extract multiple morphological features of each identified blood vessel, specifically including: The segmented scleral image patches in each quadrant are then input into the optimized YOLO object detection model. The optimized YOLO object detection model identifies all microvascular structures in the image with a width greater than or equal to 0.05 mm, and outputs the bounding box coordinates and classification confidence of each microvascular. For each identified microvessel, five morphological features are extracted from its bounding box image location: spatial orientation, vessel diameter, color value, degree of vessel tortuosity, and whether there are spots at the vessel tip. The extracted spatial orientation features are classified and assigned values ​​based on whether they point to the center of the pupil, whether they are broken, whether they point to other specific directions, or whether they are chaotic. The extracted blood vessel diameter, color value, tortuosity, and terminal spot features are quantified and graded according to a preset threshold range.

5. The scleral image recognition and analysis method based on multimodal features according to claim 4, characterized in that, The process involves matching data containing vascular morphological features and location codes with a pre-set structured scoring knowledge base to calculate the health score deduction value corresponding to each location. Specifically, this includes: A structured scoring knowledge base containing multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights is pre-built; The spatial orientation level, vessel diameter level, color value level, tortuosity level, and terminal spot level of the vascular morphology features are compared item by item with the feature entries in the structured scoring knowledge base. When all feature dimensions match the rule entries in the knowledge base, the weighted deduction is calculated for the currently identified blood vessel based on the preset weight coefficients in the rule entries. The location code of the blood vessel is compared with the feature entries in the structured scoring knowledge base item by item. When an entry in the knowledge base is matched, the weighted score is calculated for the currently identified blood vessel location (corresponding to different organs) according to the preset weight coefficient in the rule entry. Traverse all identified blood vessels within a specific location, accumulate the weighted deduction sum of all blood vessels within the specific location, and use it as the health score deduction value corresponding to the specific location; The health score deduction value corresponding to each location is bound to its location code to form a location-deduction value pair.

6. The scleral image recognition and analysis method based on multimodal features according to claim 5, characterized in that, A comprehensive analysis of the health score deduction values ​​is performed to generate a structured health assessment report that includes visual annotation charts and quantitative scores, including: The system receives the calculated health score deduction values ​​for each location, combines them with the examinee's basic physiological parameters, and calls upon a personalized correction model to correct the scores. Based on the corrected location scores, the location scores are converted into corresponding human organ health evaluation values ​​through the organ mapping matrix; By combining the health evaluation values ​​of all organs, the overall health score of the nine major systems of the human body is calculated. Integrate organ health assessment values, systemic comprehensive health scores, and basic physiological parameters to generate a structured health assessment report that includes visual annotation charts and quantitative scores; The calculated health score deduction values ​​for each location are then combined with the examinee's basic physiological parameters to apply a personalized correction model to correct the scores. Specifically, this includes: The body mass index is calculated by obtaining the age, gender, height, and weight of the person being tested. The age, gender, and body mass index are used as inputs to feed into the personalized correction model. The personalized correction model incorporates baseline parameters based on population statistics, adjusts the age-related decay factor according to the input age value, selects the gender-specific coefficient based on gender information, and compensates for the intraocular pressure impact factor based on body mass index. The personalized correction model outputs age correction coefficients, gender correction coefficients, and body mass index correction coefficients for the test subject. The health score deduction for each location is multiplied by the age correction coefficient, gender correction coefficient, and body mass index correction coefficient to obtain the final corrected health score for each location.

7. The scleral image recognition and analysis method based on multimodal features according to claim 6, characterized in that, Based on the corrected location scores, the location scores are converted into corresponding human organ health evaluation values ​​through an organ mapping matrix, specifically including: Load a preset organ mapping matrix, which defines the correspondence between each quadrant of the sclera position and the clock position, and the specific organs in the nine major systems of the human body: digestive system, nervous system, respiratory system, cardiovascular system, musculoskeletal system, endocrine system, urinary system, male reproductive system, and female reproductive system. Based on the location code in the location-deduction value pair, query one or more target human organs associated with it in the organ mapping matrix; The final location health deduction value is allocated to each associated target human organ according to a preset mapping weight. For each target human organ, the total deduction value assigned to the target human organ from all associated locations is summed. The final health evaluation value of the target human organ is obtained by subtracting the sum of the deduction values ​​from the preset basic health score of the target human organ.

8. The scleral image recognition and analysis method based on multimodal features according to claim 7, characterized in that, The comprehensive health assessment values ​​of all organs are used to calculate the overall health score of the nine major systems of the human body, specifically including: Based on the classification of human anatomy and physiology, all human organs are divided into nine major body systems; For each body system, obtain the final health assessment value of all constituent organs within that body system; Calculate the arithmetic mean of the final health evaluation values ​​of all organs under the body system, and use it as the initial comprehensive score of the body system; Based on the differences in the importance of different body systems, a preset system weight coefficient is assigned to each system. The initial comprehensive score of each body system is multiplied by its corresponding system weight coefficient, and then normalized to obtain the final comprehensive health score of the body system.

9. The scleral image recognition and analysis method based on multimodal features according to claim 8, characterized in that, The integrated organ health assessment values, system comprehensive health scores, and basic physiological parameters generate a structured health assessment report that includes visual annotation charts and quantitative scores, specifically including: Create a structured health assessment report framework, which includes a report header, core conclusions area, test details area, organ health score area, body system assessment area, health recommendations area, and report footer; In the core conclusions section, enter the overall health score and health status level calculated based on the final comprehensive health score of the body system; Embed a multi-view standard image of eye analysis with vascular feature annotation in the detection details area; The organ health scoring area lists all assessed organs and their corresponding final health assessment values ​​and risk levels in tabular form. The final comprehensive health scores and comparative analysis of the nine major systems of the human body are displayed in chart form in the body system assessment area. In the health advice area, based on the organs and systems whose risk levels exceed the threshold, the corresponding standardized health advice text is retrieved from the preset knowledge base; Enter the test subject's basic physiological parameters, test date, and unique report number into the report header, and attach a standard disclaimer at the bottom of the report to complete the report generation.

10. The scleral image recognition and analysis method based on multimodal features according to claim 9, characterized in that, The pre-built structured scoring knowledge base, which includes multi-dimensional vascular features, feature levels, corresponding organs, and deduction weights, specifically includes: Collect and organize Zhuang medicine eye diagnosis theory and clinical observation data, and classify the five dimensions of blood vessel spatial orientation, blood vessel diameter, color, degree of tortuosity, and terminal spots into three or four abnormal levels respectively; Each anomaly level feature is assigned a quantified feature code, and a base deduction value is set for features of different levels. Establish the correlation between feature coding, basic deduction values, and specific quadrants and clock positions of the sclera; Establish mapping relationships between specific quadrants of the sclera, clock positions, and human organs; Feature encoding, basic deduction values, quadrant point associations, organ mapping associations, and disease weight coefficients for specific organs are integrated into structured JSON format knowledge entries and stored as a structured scoring knowledge base file.

Images