Multi-dimensional data-based vision risk prediction model training method and system, and medium

Abstract

The invention relates to the technical field of medical images, in particular to a training method and system of a vision risk prediction model based on multi-dimensional data and a medium. The method comprises the following steps: acquiring ophthalmology historical cases, and performing retinal fluorescence angiography feature extraction according to the ophthalmology historical cases so as to obtain a retinal fluorescence angiography; carrying out retinal vessel anomaly analysis according to the retinal fluorescence angiogram, so as to obtain retinal vessel anomaly data; constructing a retina anomaly recognition model according to the retinal vessel anomaly data so as to obtain the retina anomaly recognition model; according to the retina anomaly recognition model, patient retina anomaly recognition is carried out on the ophthalmology historical cases, and therefore patient retina anomaly data are obtained; and according to the patient retina anomaly data, carrying out patient physiological test feature extraction to obtain patient physiological test data. According to the invention, more accurate vision risk prediction is realized based on the medical image technology.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a training method, system and medium for a vision risk prediction model based on multidimensional data. Background Art

[0002] Traditional vision risk prediction methods usually rely on empirical data and a single data source (such as vision testing, fundus examination results, etc.). Their prediction accuracy is limited and they cannot fully tap into potential and complex risk factors. The current vision risk prediction model based on a single data source has not been able to fully tap the potential of multidimensional data due to the single data type and limited processing methods. Traditional methods are difficult to effectively process and integrate data from multiple sources, such as imaging data, clinical diagnosis results, and physiological data. In traditional vision risk prediction models, the feature extraction process relies on manually designed rules and methods. For retinal image data, traditional methods usually use simple feature extraction methods based on color, shape, or texture, while ignoring the potential complex features and deep structural information in the image. Traditional methods have weak processing and integration capabilities for other types of data (such as genetic information, patient physiological parameters, etc.), and cannot effectively mine meaningful correlation features from high-dimensional data. This simple feature extraction method often cannot accurately capture pathological information when facing complex ophthalmic diseases, thus affecting the accuracy of prediction results. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a training method, system and medium for a vision risk prediction model based on multidimensional data to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a training method for a vision risk prediction model based on multidimensional data comprises the following steps:

[0005] Step S1: Obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; performing retinal vascular abnormality analysis according to the retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data;

[0006] Step S2: constructing a retinal abnormality recognition model based on the retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; and performing retinal abnormality recognition on historical ophthalmology cases according to the retinal abnormality recognition model, thereby obtaining retinal abnormality data of the patients;

[0007] Step S3: extracting physiological test features of the patient according to the patient's retinal abnormality data, thereby obtaining the patient's physiological test data; performing fundus image risk assessment according to the patient's physiological test data, thereby obtaining fundus image risk data;

[0008] Step S4: extracting optical coherence tomography features according to the patient's retinal abnormality data to obtain an optical coherence tomography image, and performing retinal thickness risk assessment according to the optical coherence tomography image to obtain retinal thickness risk data;

[0009] Step S5: Perform vision risk prediction model training based on the retinal thickness risk data and the fundus image risk data to obtain a vision risk prediction training model.

[0010] The present invention can accurately identify abnormalities related to retinal vascular health by extracting retinal fluorescence angiography features from historical ophthalmological cases and performing vascular abnormality analysis. This abnormal data provides an important basis for subsequent retinal disease warning. Compared with traditional methods, the present invention does not rely on empirical diagnosis, but performs risk assessment in a data-driven manner, which can discover subtle changes missed by traditional methods. The constructed retinal abnormality recognition model can automatically identify the patient's retinal abnormalities. This process greatly improves the efficiency and accuracy of recognition and avoids subjective bias and errors caused by manual analysis. Through the model, the patient's retinal abnormality data is deeply analyzed to fully capture potential pathological changes. Risk assessment combined with the patient's physiological test data and fundus images can further deepen the understanding of the patient's eye health status. Through a comprehensive assessment of the patient's physiological characteristics and fundus image risks, the model can more accurately determine the potential vision risks and provide a more comprehensive risk warning and management plan. In this process, the system can not only evaluate a single risk factor, but also comprehensively analyze multiple factors, avoiding the prediction limitations caused by single data processing in traditional methods. Optical coherence tomography feature extraction and retinal thickness risk assessment further enrich the data input of the model and increase the sensitivity to changes in retinal structure, especially in the early stages of the disease. By evaluating retinal thickness, signs of eye diseases such as glaucoma and retinopathy can be identified earlier. This process can capture subtle changes that traditional methods cannot detect, improving the accuracy of risk prediction. The vision risk prediction model is trained by integrating data from multiple dimensions, so that the model can comprehensively assess the patient's vision health risks from multiple aspects. Compared with traditional methods, the training method of the present invention can process and fuse information from different data sources, improve the adaptability and generalization ability of the model, and enable it to maintain a high prediction accuracy when facing different patient groups and different pathological conditions.

[0011] Optionally, step S1 specifically includes:

[0012] Step S11: obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images;

[0013] Step S12: performing vascular leakage analysis on the retinal fluorescence angiography image to obtain abnormal vascular leakage data;

[0014] Step S13: performing vascular occlusion analysis on the retinal fluorescence angiography image to obtain abnormal vascular occlusion data;

[0015] Step S14: performing retinal vascular abnormality integration according to the vascular leakage abnormality data and the vascular occlusion abnormality data, thereby obtaining retinal vascular abnormality data.

[0016] The vision risk prediction method proposed in the present invention significantly improves the ability to predict the risk of eye diseases by comprehensively analyzing the multidimensional data in the ophthalmological history cases, especially the feature extraction and analysis of the retinal fluorescence angiography. By extracting the key information in the retinal fluorescence angiography, the abnormality of the retinal blood vessels can be accurately identified. Unlike the limitations of the traditional method that relies on a single data source and manual feature extraction, the method of the present invention effectively overcomes the problem of neglecting the complex features of the image in the traditional method through automated data processing and analysis. In particular, for the analysis of vascular leakage and vascular occlusion of retinal fluorescence angiography, the present invention not only identifies the early signs of lesions such as vascular leakage and vascular occlusion through meticulous image analysis, but also integrates the data of the two to form more comprehensive and accurate retinal vascular abnormality data. These data can fully reflect the health status of retinal blood vessels and provide more reliable input information for subsequent vision risk prediction. Through this multi-level and multi-dimensional data processing, the model can more accurately capture potential eye health risks, especially in the early stages, it can discover abnormal changes that are difficult to identify by traditional methods. This comprehensive retinal vascular abnormality analysis method not only improves the accuracy of risk identification, but also enhances the ability to fuse multi-source data, thereby providing more scientific and accurate vision health predictions in the context of multi-dimensional data.

[0017] Optionally, step S12 is specifically:

[0018] Step S121: performing Gaussian filtering and smoothing on the retinal fluorescence angiography image, thereby obtaining a retinal fluorescence angiography smoothed image;

[0019] Step S122: Calculate the horizontal gradient according to the retinal fluorescence angiography smoothing image, so as to obtain horizontal gradient data;

[0020] Step S123: performing vertical gradient calculation according to the retinal fluorescence angiography smoothing image, thereby obtaining vertical gradient data;

[0021] Step S124: performing gradient amplitude calculation according to the horizontal gradient data and the vertical gradient data to obtain gradient amplitude data, and performing gradient amplitude map drawing according to the gradient amplitude data to obtain a gradient amplitude map;

[0022] Step S125: performing blood vessel edge detection according to the gradient amplitude map, thereby obtaining blood vessel edge data;

[0023] Step S126: performing vascular leakage analysis on the vascular edge data to obtain vascular leakage abnormality data.

[0024] The present invention provides an efficient retinal vascular abnormality detection method through an accurate retinal fluorescence angiography analysis process, which can significantly improve the accuracy of vision risk prediction. Through Gaussian filtering and smoothing, the original retinal fluorescence angiography is first denoised and smoothed, so that the noise and interference in the image are effectively suppressed, and the important structural information in the image is retained. This process optimizes the quality of subsequent feature extraction and avoids the influence of noise on data analysis. Next, by calculating the horizontal and vertical gradients of the retinal image, the edge information in the image can be extracted, and this process is crucial for identifying vascular contours and vascular abnormalities. The horizontal and vertical gradients reveal the intensity changes in the horizontal and vertical directions of the image, respectively. Combined with the calculated gradient amplitude data, the subtle structural differences in the retinal image can be clearly depicted, especially the details in the vascular area, thereby providing a solid foundation for vascular abnormality analysis. After the gradient amplitude map is drawn, the vascular edge in the retinal image can be accurately located by vascular edge detection technology, which is of great significance for further analyzing the health status of the blood vessels, especially identifying leakage or blockage abnormalities. Through accurate vascular edge detection, the present invention can efficiently extract the vascular leakage area and provide accurate data support for subsequent retinal vascular abnormality analysis. This process not only improves the accuracy of vascular abnormality detection, but also enhances the degree of automation of image processing, avoiding the subjectivity and limitations of traditional methods that rely on manual analysis. Overall, it can more comprehensively and accurately identify retinal vascular abnormalities, thereby providing more reliable and detailed data for vision risk prediction, significantly improving the accuracy and practicality of the prediction.

[0025] Optionally, step S126 is specifically:

[0026] Perform fluorescence intensity regional statistics on the blood vessel edge data to obtain high-intensity fluorescence area data;

[0027] Performing a blood vessel expansion operation according to the high-intensity fluorescence region data, thereby obtaining blood vessel expansion data;

[0028] Performing a blood vessel corrosion operation according to the blood vessel expansion data, thereby obtaining blood vessel corrosion data;

[0029] Performing vascular expansion area identification on vascular corrosion data to obtain vascular expansion area data;

[0030] Perform crack area identification on the vascular corrosion data to obtain vascular crack area data;

[0031] The leakage anomaly is integrated according to the blood vessel expansion area data and the blood vessel crack area data, so as to obtain the blood vessel leakage anomaly data.

[0032] The present invention realizes more accurate vascular leakage anomaly detection through a series of fine processing steps for vascular edge data, and significantly improves the accuracy and reliability of vision risk prediction. Fluorescence intensity regional statistics of vascular edge data can identify areas with strong fluorescence signals in the image, which usually represent abnormal areas or lesion areas of the blood vessels. By extracting high-intensity fluorescence area data, potential vascular anomalies in retinal images can be effectively identified, providing reliable basic data for subsequent analysis. Then, the high-intensity fluorescence area is processed by vascular expansion operation, so that small blood vessels or slightly abnormal areas can be further enlarged, enhancing the detection ability of vascular structural changes. The vascular expansion data provides the necessary structural basis for the subsequent vascular corrosion operation, making the detection of the entire vascular area more comprehensive and able to capture missed details. The vascular corrosion operation further refines the analysis of the vascular area, removes unnecessary noise and irrelevant areas, helps to highlight the lesion area, and has significant advantages in the identification of leakage and obstruction areas. By performing vascular expansion area identification on vascular corrosion data, the healthy part and the lesion part of the blood vessel can be effectively distinguished, and the expansion area can be accurately located, so as to better identify lesions or abnormalities. The identification of crack areas further improves the ability to detect vascular cracks and minor injuries, which are usually precursors to leakage abnormalities. Integrating the data of vascular expansion areas and vascular crack areas can comprehensively analyze the overall situation of vascular leakage abnormalities, further improving the accuracy of leakage abnormality detection. Compared with traditional methods, it not only achieves high-precision detection of vascular abnormalities, but also comprehensively improves the ability to identify leakage abnormalities through multi-level and multi-dimensional data processing. Through this series of processing, the present invention can deeply analyze the structural changes of retinal blood vessels, accurately identify the lesion area, and greatly enhance the reliability and practicality of the vision risk prediction model.

[0033] Optionally, step S13 is specifically:

[0034] Step S131: extracting retinal vascular area features from the retinal fluorescence angiography image, thereby obtaining retinal vascular area data;

[0035] Step S132: calculating the blood vessel diameter according to the retinal blood vessel area data, thereby obtaining blood vessel diameter data;

[0036] Step S133: marking the stenosis area of ​​the blood vessel diameter data, thereby obtaining the blood vessel stenosis area data;

[0037] Step S134: Counting the fluorescence intensity of the vascular stenosis area data to obtain the fluorescence intensity data of the vascular stenosis area;

[0038] Step S135: performing time change rate statistics based on the fluorescence intensity data of the vascular stenosis area, so as to obtain the vascular stenosis area with low change fluorescence intensity;

[0039] Step S136: performing optical flow calculation based on the vascular stenosis area data, thereby obtaining the vascular stenosis area optical flow data;

[0040] Step S137: extracting low optical flow vascular stenosis area features from the vascular stenosis area optical flow data, thereby obtaining low optical flow vascular stenosis area data;

[0041] Step S138: performing a vascular occlusion anomaly intersection operation based on the low optical flow vascular stenosis area data and the low variation fluorescence intensity vascular stenosis area data, thereby obtaining vascular occlusion anomaly data.

[0042] The present invention can accurately locate the vascular area from the image and obtain the spatial distribution and morphological characteristics of the blood vessels by extracting the vascular area features from the retinal fluorescence angiography. This step provides key structural information for subsequent vascular analysis. Next, by calculating the vascular diameter, the width of the blood vessel can be quantified, thereby identifying abnormal vascular areas, especially the narrowed vascular parts, and further providing an important basis for the assessment of vascular health status. After marking the narrowed area of ​​the vascular diameter data, it can be clearly determined which blood vessels are narrowed, and further provide early warning for vascular obstruction or other lesions. The fluorescence intensity statistics of the vascular stenosis area can reveal the pathological changes inside and around the blood vessels, especially the fluorescence intensity in the vascular stenosis area is often related to the changes in blood flow and the degree of lesions. Through the time change rate statistics of the fluorescence intensity data, the present invention can further distinguish which narrowed areas are stable and which areas have a trend of change. This analysis helps to identify potential vascular lesions, especially in the early stages of vascular stenosis. Further, by optical flow calculation, the dynamic changes of blood flow in the vascular stenosis area are modeled, and the speed and direction changes of blood flow can be quantified, so as to better understand the flow conditions in the blood vessels. Feature extraction of low optical flow vascular stenosis areas helps to find areas with slow blood flow, which are precursors to vascular obstruction. Intersection operations are performed on low optical flow vascular stenosis area data and low variable fluorescence intensity vascular stenosis areas to comprehensively analyze vascular obstruction abnormalities and further improve the accuracy of detecting obstruction abnormalities. It is possible to comprehensively analyze changes in retinal blood vessels from multiple dimensions and levels, accurately identify pathological features such as vascular stenosis and obstruction, provide more accurate data support for vision risk prediction models, and effectively improve the accuracy and reliability of prediction results. Compared with traditional methods, the present invention can reveal potential risks to eye health more comprehensively and deeply through automated and refined feature extraction and analysis, and has significant innovation and practicality.

[0043] Optionally, step S4 is specifically:

[0044] Step S41: extracting optical coherence tomography features according to the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image;

[0045] Step S42: performing retinal layer segmentation on the optical coherence tomography image, thereby obtaining retinal nerve fiber layer data and retinal inner nuclear layer data;

[0046] Step S43: performing thickness time series risk analysis on the retinal nerve fiber layer data, thereby obtaining nerve fiber layer thickness time risk data;

[0047] Step S44: performing regional thickness difference risk analysis on the retinal inner nuclear layer data, thereby obtaining regional thickness difference risk data of the inner nuclear layer;

[0048] Step S45: integrating the retinal thickness risk data according to the nerve fiber layer thickness time risk data and the inner nuclear layer regional thickness difference risk data, thereby obtaining the retinal thickness risk data.

[0049] The present invention accurately obtains the details of the retinal layer through optical coherence tomography feature extraction, laying the foundation for further analysis. Through retinal layer segmentation, the data of the retinal nerve fiber layer and the inner nuclear layer can be refined, and the pathological information of different layers can be distinguished, which helps to better understand the health status of each retinal layer. The time series risk analysis of nerve fiber layer thickness can track the changes in the nerve fiber layer, timely discover the trend of thickness changes over time, and identify early signals of pathological changes, thereby providing decision support for disease prevention and intervention. At the same time, the risk analysis of the difference in thickness of the inner nuclear layer further analyzes the regional differences in the inner nuclear layer of the retina, identifies the existing local thickness changes, and thus provides a more detailed risk assessment, which is especially important for early warning of retinal abnormalities in the early stage. Retinal thickness risk integration combined with the time risk data of the thickness of the nerve fiber layer and the risk data of the difference in thickness of the inner nuclear layer can comprehensively and systematically evaluate the overall health status of the retina. This integration process effectively integrates retinal data from different levels, eliminates the limitations of traditional methods on a single data source, and can comprehensively consider the changing trends of each layer of the retina, providing more in-depth and broad support for vision risk prediction. Compared with the simple feature extraction of traditional methods, comprehensive analysis of multidimensional data can better reveal the complexity of retinal pathological changes, improve the accuracy of prediction, help to conduct risk assessment of retinal diseases earlier and more accurately, and enhance the scientificity and reliability of clinical diagnosis.

[0050] Optionally, step S43 is specifically:

[0051] Step S431: Counting the thickness change of the retinal nerve fiber layer data, thereby obtaining the thickness change data of the retinal nerve fiber layer;

[0052] Step S432: calculating the change rate according to the retinal nerve fiber layer thickness change data, thereby obtaining thickness change rate data;

[0053] Step S433: obtaining thickness change rate risk threshold data;

[0054] Step S434: performing risk classification on the thickness change rate data according to the thickness change rate risk threshold data, thereby obtaining nerve fiber layer thickness time risk data.

[0055] The present invention helps to reveal the changing trend of retinal thickness at each time point by statistically analyzing the thickness changes of the retinal nerve fiber layer data, and provides basic data for subsequent analysis. By calculating the thickness change rate, the speed of the change in the thickness of the retinal nerve fiber layer can be quantified. This process can help identify early signs of retinal disease, especially in the stage where the disease changes more slowly. Obtaining the thickness change rate risk threshold data provides a standardized reference value for risk judgment, which can effectively distinguish between normal and abnormal ranges of change, thereby ensuring the accuracy and reliability of risk division. Finally, the thickness change rate data is divided into risks according to these thresholds, and the thickness changes of the retinal nerve fiber layer of different patients can be accurately stratified to form nerve fiber layer thickness time risk data. Compared with the single data source analysis of the traditional method, the present invention provides a multi-dimensional, dynamic and accurate risk assessment method, which can give full play to the potential of various types of data and improve the accuracy of prediction.

[0056] Optionally, step S44 is specifically:

[0057] Step S441: dividing the retinal inner nuclear layer data into regions, thereby obtaining the inner nuclear layer central region data and the inner nuclear layer peripheral region data;

[0058] Step S442: Calculating the average thickness according to the data of the central region of the inner core layer, thereby obtaining the average thickness data of the central region of the inner core layer;

[0059] Step S443: Calculating the average thickness according to the data of the peripheral area of ​​the inner core layer, thereby obtaining the average thickness data of the peripheral area of ​​the inner core layer;

[0060] Step S444: performing thickness difference significance judgment based on the average thickness data of the central region of the inner core layer and the average thickness data of the peripheral region of the inner core layer, thereby obtaining thickness difference significance data;

[0061] Step S445: Perform risk assessment based on the thickness difference significance data to obtain inner core layer region thickness difference risk data.

[0062] The present invention can analyze different areas of the retina more carefully by dividing the data of the inner nuclear layer of the retina, and provide accurate basic data for the subsequent thickness difference assessment. By calculating the average thickness of the central area and the peripheral area of ​​the inner nuclear layer respectively, more detailed retinal hierarchical structure information can be obtained, which is crucial for analyzing the health status of the retina. Then, the thickness difference significance is judged according to the average thickness data of the two areas, and whether the thickness difference between the central area and the peripheral area is statistically significant is effectively identified, thereby providing a scientific basis for potential pathological changes. This process helps to exclude thickness changes caused by normal physiological differences and ensures the accuracy of the assessment. Finally, the risk assessment based on the significance data not only reveals the clinical significance of the inner nuclear layer thickness difference, but also can further improve the sensitivity and specificity of vision risk prediction. This method overcomes the limitations of traditional methods for data processing and analysis, can comprehensively consider the thickness differences in different areas, more comprehensively assess the health status of the retina, and provide more reliable risk assessment results.

[0063] Optionally, the present specification also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the computer program implements any one of the methods for training a vision risk prediction model based on multidimensional data.

[0064] Optionally, the present specification also provides a training system for a vision risk prediction model based on multidimensional data, which is used to execute the training method for a vision risk prediction model based on multidimensional data as described above. The training system for a vision risk prediction model based on multidimensional data includes:

[0065] Retinal vascular abnormality analysis module: used to obtain historical ophthalmological cases, and extract features of retinal fluorescence angiography images based on the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; perform retinal vascular abnormality analysis based on retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data;

[0066] Retinal abnormality recognition model construction module: used to construct a retinal abnormality recognition model based on retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; identify retinal abnormalities of patients in ophthalmic history cases based on the retinal abnormality recognition model, thereby obtaining retinal abnormality data of patients;

[0067] Fundus image risk assessment module: used to extract physiological test features of patients based on abnormal retinal data of patients, so as to obtain physiological test data of patients; perform fundus image risk assessment based on physiological test data of patients, so as to obtain fundus image risk data;

[0068] Retinal thickness risk assessment module: used to extract optical coherence tomography features based on the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image, and to perform retinal thickness risk assessment based on the optical coherence tomography image, thereby obtaining retinal thickness risk data;

[0069] Vision risk prediction model training module: used to train the vision risk prediction model based on retinal thickness risk data and fundus image risk data, so as to obtain a vision risk prediction training model.

[0070] The present invention provides a training system for a vision risk prediction model based on multidimensional data. The system can implement any training method for a vision risk prediction model based on multidimensional data of the present invention. The system is used to combine the operations between various modules and a medium for signal transmission to complete the training method for a vision risk prediction model based on multidimensional data. The internal modules of the system cooperate with each other to achieve more accurate vision risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0072] Figure 1 It is a schematic flow chart of the steps of the training method of the vision risk prediction model based on multidimensional data of the present invention;

[0073] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0074] Figure 3 Detailed step flow diagram of step S12 in the present invention;

[0075] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0076] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0077] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0078] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0079] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for training a vision risk prediction model based on multidimensional data, the method comprising the following steps:

[0080] Step S1: Obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; performing retinal vascular abnormality analysis according to the retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data;

[0081] In this embodiment, it is necessary to obtain multiple ophthalmic historical case data, which generally include the patient's retinal fluorescence angiography and related clinical diagnosis information. After obtaining the retinal fluorescence angiography, the image is preprocessed, such as denoising, contrast enhancement, etc., so as to extract important structural information in the image. During image processing, the image is smoothed using a Gaussian filter to reduce the impact of noise in the image, and then an edge detection algorithm (such as the Canny algorithm) is applied to extract the boundaries of the blood vessels. For the abnormal analysis of retinal blood vessels, the image segmentation technology is used to distinguish the retinal blood vessel area from other areas, and the characteristics of vascular leakage and vascular occlusion are combined to calculate the diameter change of the blood vessel, the morphological change of the blood vessel, etc., and the data of the abnormal blood vessel is extracted. A further step includes segmenting the image based on a set threshold (such as a fluorescence intensity threshold, a blood vessel diameter change threshold, etc.), extracting the data of the abnormal blood vessel area, and these data provide a basis for the construction of the subsequent retinal abnormality recognition model.

[0082] Step S2: constructing a retinal abnormality recognition model based on the retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; and performing retinal abnormality recognition on historical ophthalmology cases according to the retinal abnormality recognition model, thereby obtaining retinal abnormality data of the patients;

[0083] In this embodiment, based on the retinal vascular abnormality data, a series of feature extraction methods are used to further analyze the abnormal data. For the vascular abnormality area, structural features (such as vascular curvature, vascular branching angle, vascular density, etc.) and optical features (such as fluorescence intensity, vascular filling rate, etc.) are used to build a model. Then, these features are used to construct a retinal abnormality recognition model, and a threshold judgment method is used to distinguish the vascular abnormality area from the normal area. During the model construction process, a specific abnormal threshold value can be set (such as a change in vascular diameter greater than 0.5mm, a change in fluorescence intensity greater than 5%, etc.), and these thresholds can be used to classify the data to determine whether there is an abnormality in the retinal blood vessels. Through training on a large number of historical ophthalmic cases, a retinal abnormality recognition model is formed, so that the model can automatically identify the abnormality of the patient's retina based on the input vascular data, and output the corresponding retinal abnormality data.

[0084] Step S3: extracting physiological test features of the patient according to the patient's retinal abnormality data, thereby obtaining the patient's physiological test data; performing fundus image risk assessment according to the patient's physiological test data, thereby obtaining fundus image risk data;

[0085] In this embodiment, physiological test features are extracted using the patient's retinal abnormality data. These physiological test features include, but are not limited to, the patient's blood sugar level, intraocular pressure, speed of vision loss, genetic information, etc., and specific features can be obtained through medical equipment or clinical tests. Next, by performing a risk assessment on the patient's fundus image, an image-based analysis method is used, such as calculating the average density of retinal blood vessels, the uniformity of blood vessel distribution and other parameters, and a comprehensive assessment is performed in combination with the patient's physiological test data. A risk assessment model is set according to existing standards or clinical research data, and the physiological features are merged with the image features using the weighted average method to ultimately obtain the patient's fundus image risk data. The key to this step is the joint analysis of the structural information and physiological information in the fundus image to obtain the ability to predict abnormal retinal changes.

[0086] Step S4: extracting optical coherence tomography features according to the patient's retinal abnormality data to obtain an optical coherence tomography image, and performing retinal thickness risk assessment according to the optical coherence tomography image to obtain retinal thickness risk data;

[0087] In this embodiment, optical coherence tomography (OCT images) are extracted from the patient's retinal abnormality data, and these images are processed, and the retinal layers (such as the retinal nerve fiber layer, inner nuclear layer, etc.) are accurately segmented using a segmentation algorithm. Common segmentation methods include threshold-based segmentation, region growing algorithm, level set model, etc. These methods can effectively extract the thickness data of each layer, and then perform a risk assessment of thickness changes. For the retinal nerve fiber layer, by calculating the thickness changes of the layer at different time points, the time series data of the thickness is obtained, and the rate of change of these thickness data is calculated. By setting a certain rate of change threshold (for example, the thickness change rate exceeds a certain set value, such as 0.3% / year), the risk assessment of retinal thickness is performed. For the inner nuclear layer of the retina, a similar method is used to perform regional thickness difference analysis, and the significance of the thickness difference between each region is calculated. The purpose of these steps is to provide a basis for subsequent vision risk prediction through the law of retinal thickness changes.

[0088] Step S5: Perform vision risk prediction model training based on the retinal thickness risk data and the fundus image risk data to obtain a vision risk prediction training model.

[0089] In this embodiment, retinal thickness risk data and fundus image risk data of a large number of patients are collected as input features as the basis for model training. For model training, traditional machine learning algorithms such as decision trees, random forests, support vector machines (SVM) can be used, or deep learning methods such as convolutional neural networks (CNN) can be used for training according to the complexity of the data. During the training process, the model performance is optimized by adjusting model parameters (such as tree depth, learning rate, number of iterations, etc.). The trained model can predict the vision risk based on the new patient data and output the corresponding risk level or probability of decreased vision. The key to this prediction model is how to combine the comprehensive information of retinal thickness data and fundus image data to improve the ability to predict changes in vision.

[0090] Optionally, step S1 specifically includes:

[0091] Step S11: obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images;

[0092] In this embodiment, the patient's retinal fluorescence angiography (FFA) and related clinical diagnostic data are collected. Retinal fluorescence angiography is usually obtained by fluorescein fundus angiography (FA) technology, which uses fluorescein dye injected into the blood and uses a fundus camera to capture dynamic images of fundus blood vessels. After the image is acquired, each fluorescence angiography is preprocessed, including image denoising, illumination balancing, and image enhancement. Specifically, a Gaussian filter is first applied to smooth the image to remove noise and smooth brightness differences. Then, the image is contrast enhanced using histogram equalization technology to ensure that the edges of the blood vessels are more clearly visible. On this basis, an edge detection algorithm (such as the Canny algorithm) is used to extract the boundaries of the retinal blood vessels, thereby obtaining a clear vascular feature map. Ultimately, these vascular boundaries and image features provide basic data for subsequent vascular abnormality analysis.

[0093] Step S12: performing vascular leakage analysis on the retinal fluorescence angiography image to obtain abnormal vascular leakage data;

[0094] In this embodiment, starting from the characteristics of the vascular image, the edges and surrounding areas of the blood vessels are located. For each vascular area, a local contrast enhancement algorithm is used to extract fluorescence intensity data. These fluorescence intensity data can reflect the leakage around the blood vessels. Vascular leakage is usually manifested as an abnormal increase in fluorescence intensity. The specific operation is: setting a fluorescence intensity threshold in the image, for example, treating the position where the fluorescence intensity is greater than the set value (such as 150 units) as a potential leakage area. By marking these areas as leakage areas, preliminary leakage data is obtained. Next, the regional growing algorithm is applied to expand the leakage area, identify the scope of the leakage, and confirm the boundary of the leakage area by comparing the fluorescence intensity with the normal area. Finally, the data of all leakage areas will be integrated to form abnormal vascular leakage data, providing a basis for the next step of abnormal identification.

[0095] Step S13: performing vascular occlusion analysis on the retinal fluorescence angiography image to obtain abnormal vascular occlusion data;

[0096] In this embodiment, each blood vessel is analyzed using a blood vessel edge map, especially the diameter and morphological characteristics of the blood vessel. Vascular obstruction is usually manifested as a sudden change in the diameter of the blood vessel or a local contraction, and a diameter threshold is used to detect whether there is vascular stenosis or obstruction. For example, by calculating the average diameter of each blood vessel and setting a threshold (such as a blood vessel with a diameter less than 0.5 mm is considered to be blocked), the area of ​​stenosis or obstruction is found. Subsequently, the blocked area is processed using a local contrast enhancement algorithm and compared with the normal state of the surrounding blood vessels to detect whether there is a blood flow obstruction. For the blocked area, time series analysis is also required to track the changes in blood vessel obstruction. If an obvious blocking trend is found, the data of the blood vessel obstruction area is further extracted, and these data are integrated to form abnormal blood vessel obstruction data. These data can not only reflect the degree of blood vessel obstruction, but also provide important clues for the comprehensive analysis of retinal vascular abnormalities.

[0097] Step S14: performing retinal vascular abnormality integration according to the vascular leakage abnormality data and the vascular occlusion abnormality data, thereby obtaining retinal vascular abnormality data.

[0098] In this embodiment, the abnormal data of vascular leakage and the abnormal data of vascular occlusion are compared and integrated. The key to integration is to associate different types of abnormalities (such as leakage and occlusion) according to spatial positions. For example, a spatial adjacency standard can be set to merge adjacent leakage areas and occlusion areas to confirm whether they are abnormal manifestations on the same blood vessel. In addition, in order to improve the accuracy of integration, the weighted average method can be used to integrate the leakage and occlusion data, and the weighted coefficients of the leakage area and the occlusion area can be set to indicate the severity of the two abnormalities. A unified abnormality identification standard is set. For example, when vascular leakage and occlusion occur at the same time, the degree of abnormality is higher and a higher weight is given. Ultimately, the integrated retinal vascular abnormality data will contain information such as the abnormal type, location, severity, etc. of each blood vessel, providing complete vascular abnormality data for subsequent retinal abnormality identification and risk assessment. This integration process is completed automatically through the algorithm, minimizing manual intervention and improving the accuracy and efficiency of abnormality detection.

[0099] Optionally, step S12 is specifically:

[0100] Step S121: performing Gaussian filtering and smoothing on the retinal fluorescence angiography image, thereby obtaining a retinal fluorescence angiography smoothed image;

[0101] In this embodiment, a Gaussian filter is applied to perform convolution processing on the original image. The specific operation includes setting the standard deviation (σ) of the Gaussian kernel function. The standard deviation determines the degree of smoothing. The larger the σ value, the more significant the image smoothing effect and the more details are lost. The σ value is usually set in the range of 1.0 to 2.0, and the optimal value is determined by experiment. During the filtering process, the Gaussian kernel function performs weighted averaging on each pixel point and its neighboring pixels in the image to reduce the influence of noise and image details, thereby improving the overall smoothness of the image. The retinal fluorescence angiography after Gaussian filtering can remove noise, making subsequent edge detection and gradient calculation more accurate.

[0102] Step S122: Calculate the horizontal gradient according to the retinal fluorescence angiography smoothing image, so as to obtain horizontal gradient data;

[0103] In this embodiment, when calculating the horizontal gradient in the retinal fluorescence angiography smoothing image, the Sobel operator is first used to calculate the horizontal gradient in the image. The specific method is to perform a convolution operation on the image through a 3x3 Sobel convolution kernel. The horizontal convolution kernel of the Sobel operator is:

[0104]

[0105] The convolution kernel calculates the rate of change of each pixel and its surrounding pixels in the horizontal direction. Through this calculation, the gradient value of the image in the horizontal direction can be obtained, reflecting the brightness change in the horizontal direction of the image. The horizontal gradient data will provide important directional information for subsequent edge detection and blood vessel analysis, especially in the direction and structure recognition of blood vessels in the image.

[0106] Step S123: performing vertical gradient calculation according to the retinal fluorescence angiography smoothing image, thereby obtaining vertical gradient data;

[0107] In this embodiment, the vertical gradient calculation is similar to the horizontal gradient calculation, and the vertical convolution kernel of the Sobel operator is used to calculate the rate of change in the vertical direction of the image. The vertical convolution kernel of the Sobel operator is:

[0108]

[0109] By convolving this kernel with the retinal fluorescence angiography smoothing map, the vertical gradient value of each pixel in the image is calculated. The vertical gradient data reflects the brightness change of the image in the vertical direction and is usually used to capture the vertical edges and vascular morphology in the image. The combination of horizontal and vertical gradients helps to fully identify the edges of blood vessels and other important structures.

[0110] Step S124: performing gradient amplitude calculation according to the horizontal gradient data and the vertical gradient data to obtain gradient amplitude data, and performing gradient amplitude map drawing according to the gradient amplitude data to obtain a gradient amplitude map;

[0111] In this embodiment, after calculating the horizontal and vertical gradients, the gradient amplitude is calculated next. The calculation formula of the gradient amplitude is:

[0112]

[0113] Among them, Gx and Gy are the horizontal gradient and vertical gradient data respectively. By calculating the gradient amplitude of each pixel, an image containing the edge strength information of each point in the image can be obtained. Gradient amplitude data is usually expressed as the intensity of the edge position in the image. The gradient amplitude map is generated by plotting these amplitude data, which can effectively reveal the significant edges in the image, such as the edges of blood vessels, the boundaries of the choroid, etc. This map provides a clear image basis for subsequent blood vessel edge detection.

[0114] Step S125: performing blood vessel edge detection according to the gradient amplitude map, thereby obtaining blood vessel edge data;

[0115] In this embodiment, after obtaining the gradient amplitude map, an edge detection algorithm (such as Canny edge detection) is applied to further process the image. The Canny edge detection algorithm screens the edges in the image by setting two thresholds. First, a low threshold and a high threshold are used for double threshold processing. The low threshold is used to detect weaker edges, while the high threshold is used to retain strong edges. In the gradient amplitude map, all pixels with gradient amplitudes greater than the high threshold will be marked as edges, while pixels below the low threshold will be ignored. Pixels between the two thresholds will be retained based on their connectivity. In this process, the results of the edge detection will highlight the edges and directions of blood vessels in the retinal image to form blood vessel edge data.

[0116] Step S126: performing vascular leakage analysis on the vascular edge data to obtain vascular leakage abnormality data.

[0117] In this embodiment, the edge intensity of each blood vessel in the edge image is calculated, and the fluorescence intensity regional analysis is performed on it. By setting an intensity threshold (for example, 200 units), the area with fluorescence intensity greater than this value is detected and marked as a potential leakage area. The leakage area is usually manifested as an abnormal increase in fluorescence intensity, reflecting the abnormality of the blood vessel. Subsequently, the regional growing algorithm is applied to expand the leakage area, identify its boundaries, and compare it with other normal areas. Finally, by analyzing the difference in fluorescence intensity in different blood vessel areas, the severity and location of vascular leakage are obtained. These leakage abnormality data will provide an important basis for subsequent retinal abnormality identification.

[0118] Optionally, step S126 is specifically:

[0119] Perform fluorescence intensity regional statistics on the blood vessel edge data to obtain high-intensity fluorescence area data;

[0120] In this embodiment, on the basis of the blood vessel edge data, the fluorescence intensity of each pixel in the image is extracted. The fluorescence intensity can be represented by the gray value of each pixel in the image. Usually, in the retinal fluorescence angiography, the fluorescence intensity of the blood vessel edge is higher. In order to accurately identify the high-intensity area, it is necessary to set a threshold value (for example, the threshold value is 150 units), and only consider the area with fluorescence intensity greater than this value as the high-intensity fluorescence area. Using image processing software or programming tools (such as OpenCV), the blood vessel edge area is divided into regions, and through thresholding operation, the pixels with intensity higher than the threshold value are marked as high-intensity fluorescence areas. The total number of pixels, area, position and other parameters of these areas are counted to obtain high-intensity fluorescence area data. These data reflect the part of the blood vessel where there is leakage or lesions, and provide data support for subsequent operations.

[0121] Performing a blood vessel expansion operation according to the high-intensity fluorescence region data, thereby obtaining blood vessel expansion data;

[0122] In this embodiment, the blood vessel dilation operation is an operation in image morphological processing, which is intended to expand the boundary of the high-intensity fluorescence area to better identify vascular lesions. When using the dilation operation, a structural element is first defined, and a rectangular structural element of size 3x3 is usually selected. Then, through the dilation operation, the edge of the high-intensity fluorescence area is expanded to cover the pixels outside the original area. In this process, for each pixel, the pixels in its surrounding neighborhood will be checked, and if there are high-intensity fluorescent pixels in the neighborhood, the pixel will also be marked as the dilated area. The dilation process can be implemented using an image processing library (such as cv2.dilate in OpenCV). In the dilated area, the vascular morphology becomes more obvious, which helps to further analyze the pathological characteristics of the blood vessels. Finally, the obtained vascular dilation data is used to detect whether the blood vessels have further expansion or deformation.

[0123] Performing a blood vessel corrosion operation according to the blood vessel expansion data, thereby obtaining blood vessel corrosion data;

[0124] In this embodiment, the vascular erosion operation is another basic operation in image morphology, which is intended to eliminate the false boundaries generated during the expansion process. The erosion operation uses a structural element opposite to the expansion operation, and the purpose is to reduce the scope of the expansion area. First, a structural element is defined, such as a rectangular or circular structural element of size 3x3. Then, the expanded vascular area is eroded by comparing each pixel with the structural element in its neighborhood. If all pixels in the neighborhood are target pixels, the central pixel is retained, otherwise it will be deleted. The purpose of the erosion operation is to remove the unreal expansion part, refine the vascular boundary, and reduce noise and irregular areas. Use an image processing library (such as cv2.erode in OpenCV) to implement the erosion operation. After corrosion, the actual boundary of the blood vessel is clearer, which helps to reduce the interference of non-vascular areas and provide more accurate lesion area data.

[0125] Performing vascular expansion area identification on vascular corrosion data to obtain vascular expansion area data;

[0126] In this embodiment, when the blood vessel expansion area is identified in the data after the blood vessel is corroded, the morphological operation is first used to detect whether there is an expanded area in the image. The corroded image is processed using a region labeling algorithm (such as connected component analysis), and each connected pixel area is marked as an expansion area. Among these expansion areas, special attention is paid to those areas that are in contact with the edge area of ​​the blood vessel and are larger in size. These areas are usually manifestations of blood vessel expansion. By setting the area threshold of the area, the blood vessel expansion area that meets the conditions is screened out, such as the area greater than a certain number of pixels (such as 500 pixels). By calculating the shape characteristics of the expansion area (such as aspect ratio, circularity, etc.), the blood vessel expansion area and other types of areas are further distinguished to obtain the blood vessel expansion area data. This step is of great significance for identifying vascular lesions such as expansion or curvature.

[0127] Perform crack area identification on the vascular corrosion data to obtain vascular crack area data;

[0128] In this embodiment, crack regions appear in the image after the blood vessel corrosion, and these regions are usually manifested as gaps or separated parts in the image. Using the edge detection results of the image, the crack regions can be identified by extracting small holes or gaps in the image. First, the low-intensity regions related to the edges of the blood vessels in the image after corrosion are used to detect the crack regions in the image using connected region analysis. A suitable crack region area threshold is set (such as an area less than 200 pixels is considered a crack region), and the shape features of these regions, such as the width, length and morphology of the cracks, are calculated. The details of the crack region are further extracted using an edge enhancement algorithm to ensure that the fractures inside or between blood vessels can be accurately identified. Through this process, the data of the blood vessel crack region is obtained, which provides a basis for further analyzing the causes of vascular leakage.

[0129] The leakage anomaly is integrated according to the blood vessel expansion area data and the blood vessel crack area data, so as to obtain the blood vessel leakage anomaly data.

[0130] In this embodiment, the data of the expansion area and the crack area are combined, and the morphology, size, position and other characteristics of each area are comprehensively analyzed. By setting a leakage area identification threshold (for example, an area with an area of ​​more than 100 pixels is considered an abnormal area), the leakage areas corresponding to the vascular expansion and cracks are identified. Using the fluorescence intensity information of the area, it is further determined which expansion areas or crack areas have abnormally high intensities, thereby confirming that these areas are leakage areas. By combining the data of the vascular expansion and crack areas, a data set of leakage abnormal areas is generated, which represent vascular lesions or dysfunctions, such as vascular leakage, blood loss and other problems. Ultimately, the integrated data provides a more comprehensive leakage abnormality assessment, providing a basis for subsequent clinical analysis.

[0131] Optionally, step S13 is specifically:

[0132] Step S131: extracting retinal vascular area features from the retinal fluorescence angiography image, thereby obtaining retinal vascular area data;

[0133] In this embodiment, the retinal fluorescence angiography is preprocessed, such as denoising and contrast enhancement, so as to better distinguish the vascular area. The edge of the blood vessel is extracted using an edge detection algorithm (such as Canny edge detection). Next, morphological operations (such as expansion and corrosion) are used to close the edge to fill the holes caused by the gaps between the blood vessels, so as to better define the vascular area. Based on the morphological characteristics of the blood vessels (such as the connectivity and width of the blood vessels), the blood vessel area data is further extracted. These area data contain the distribution information of the blood vessels and can provide basic data for subsequent blood vessel diameter calculation and abnormal analysis.

[0134] Step S132: calculating the blood vessel diameter according to the retinal blood vessel area data, thereby obtaining blood vessel diameter data;

[0135] In this embodiment, the data of the blood vessel area is used to calculate the blood vessel diameter by selecting pixel points in different areas. The blood vessel is intercepted along its longitudinal direction to obtain the blood vessel width at that position, which is usually calculated on each blood vessel cross section using a scanning line algorithm. In order to obtain more accurate blood vessel diameter data, multiple cutting surfaces can be selected for averaging, or morphological methods (such as skeletonization) can be used in the image to extract the blood vessel centerline, and the blood vessel diameter is calculated through the centerline. The blood vessel diameter is usually calculated in pixels, and then the pixels are converted into physical units (such as millimeters) using a known proportionality factor. The data output by this step will reflect the diameter distribution of retinal blood vessels and provide support for marking areas of vascular stenosis.

[0136] Step S133: marking the stenosis area of ​​the blood vessel diameter data, thereby obtaining the blood vessel stenosis area data;

[0137] In this embodiment, a blood vessel diameter threshold is set (e.g., blood vessel diameters less than 50% of the average diameter value are considered stenosis areas). Then, in the blood vessel diameter data, all blood vessel parts that are less than the threshold are marked, and these parts are stenosis areas. In order to improve the accuracy of marking, a sliding window method is used to perform local analysis on the blood vessel area, and by comparing the difference between the local blood vessel diameter and the global blood vessel diameter, it is determined whether it is a stenosis area. Each area marked as stenosis contains diameter information and the degree of change relative to the surrounding blood vessels. The data set output by this step contains the location, size, and other geometric features of the stenosis area of ​​the blood vessel, providing a basis for subsequent fluorescence intensity statistics.

[0138] Step S134: Counting the fluorescence intensity of the vascular stenosis area data to obtain the fluorescence intensity data of the vascular stenosis area;

[0139] In this embodiment, in the stenotic area, the fluorescence intensity value of each pixel is extracted, and the average fluorescence intensity, maximum intensity and intensity standard deviation of the area are calculated. The statistical data of fluorescence intensity are usually affected by the vascular structure, the distribution of fluorescent dyes and the blood flow state. Therefore, by calculating these statistical values, the fluorescence intensity distribution information of the vascular stenosis area can be obtained. In this step, it is first necessary to extract the fluorescence intensity data in the stenotic area according to the set vascular stenosis threshold. Then, these data are statistically analyzed to obtain the fluorescence intensity information of each stenotic area, and then identify the vascular area with lower or higher intensity. The goal of this step is to provide more accurate intensity data for further evaluation of the blood vessel.

[0140] Step S135: performing time change rate statistics based on the fluorescence intensity data of the vascular stenosis area, so as to obtain the vascular stenosis area with low change fluorescence intensity;

[0141] In this embodiment, a time series analysis method is used to compare the fluorescence intensity of the same vascular stenosis area at different time points. For each vascular stenosis area, the fluorescence intensity at multiple sampling time points is counted, and the rate of change of the fluorescence intensity over time is calculated. Specifically, the rate of change of fluorescence intensity (for example: the amount of change in fluorescence intensity per second) can be calculated, and the vascular areas can be classified according to the magnitude of the change rate. If the rate of change of fluorescence intensity is low, it is considered that the state of vascular stenosis in this area is relatively stable and may be a chronic lesion; if the rate of change is high, it indicates that there is a risk of acute or rapid changes in vascular stenosis in this area. This process provides dynamic change characteristics of vascular stenosis areas, and further provides information for vascular health assessment.

[0142] Step S136: performing optical flow calculation based on the vascular stenosis area data, thereby obtaining the vascular stenosis area optical flow data;

[0143] In this embodiment, optical flow calculation is used to analyze the dynamic movement of the vascular stenosis area in the retinal image. By processing the continuous images of the vascular area, the displacement and speed of the pixels in the blood vessel are calculated using the optical flow algorithm. The image of the vascular stenosis area is extracted from the retinal fluorescence angiography, and the two consecutive frames of images are compared using the optical flow calculation method (such as the Horn-Schunck algorithm or the Lucas-Kanade method) to calculate the movement direction and speed of each pixel in the vascular stenosis area. This method captures the dynamic changes of the vascular area by estimating the movement of pixels, especially whether the vascular stenosis part has a large deformation or movement in a short period of time. According to the optical flow data, the movement characteristics of the blood vessels can be further identified, and provide an important basis for subsequent abnormal diagnosis.

[0144] Step S137: extracting low optical flow vascular stenosis area features from the vascular stenosis area optical flow data, thereby obtaining low optical flow vascular stenosis area data;

[0145] In this embodiment, after obtaining the optical flow data of the vascular stenosis area, it is necessary to extract the vascular stenosis area with low optical flow. These areas are usually manifestations of vascular stenosis or blockage. By setting the optical flow rate threshold (for example: the area below a certain optical flow rate value is regarded as a low optical flow area), those areas with low optical flow rates are marked, which represent the low blood flow state of the blood vessel. By analyzing the geometric features of the low optical flow area (such as shape, size, rate of change, etc.), it is further identified whether these areas have potential vascular blockage or stenosis problems. The key to this step is to extract those narrow areas that show slow or no blood flow, and provide data support for the detection of vascular obstruction.

[0146] Step S138: performing a vascular occlusion anomaly intersection operation based on the low optical flow vascular stenosis area data and the low variation fluorescence intensity vascular stenosis area data, thereby obtaining vascular occlusion anomaly data.

[0147] In this embodiment, the vascular stenosis areas with low optical flow and low variable fluorescence intensity are marked respectively, and the overlapping areas of the two are found through intersection operation. These overlapping areas usually represent more serious blockage of blood vessels. By calculating the geometric characteristics of these areas (such as area, shape, distribution, etc.), it is possible to further confirm which areas of vascular stenosis have caused partial blockage of blood vessels or poor blood flow. Ultimately, the intersection data provides accurate positioning and analysis basis for the identification of abnormal vascular obstruction.

[0148] Optionally, step S4 is specifically:

[0149] Step S41: extracting optical coherence tomography features according to the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image;

[0150] In this embodiment, the OCT image is preprocessed, such as denoising, contrast enhancement and equalization, to ensure that the image quality meets the analysis requirements. Then, an edge detection algorithm (such as Canny edge detection) is applied to identify the boundaries of different levels of the retina, especially to identify important structures such as the nerve fiber layer, inner nuclear layer and retinal blood vessels. During the feature extraction process, the variation curve of the optical reflection intensity is extracted, and the thickness information and its changes at different levels in the image are analyzed. Feature extraction also includes geometric morphological analysis of the retina, such as data such as curvature changes and the distance between layers. These feature data provide a basis for subsequent risk analysis. Ultimately, the extracted image data can describe the structural changes of each layer of the retina, providing a basis for further analysis of the thickness of the retina, interlayer differences, etc.

[0151] Step S42: performing retinal layer segmentation on the optical coherence tomography image, thereby obtaining retinal nerve fiber layer data and retinal inner nuclear layer data;

[0152] In the present embodiment, after obtaining the optical coherence tomography image, the segmentation of the retinal layer is performed, and the main purpose is to clearly divide the different levels (such as nerve fiber layer, inner core layer, etc.) in the retinal image. The brightness value in the image is processed using a threshold segmentation method, and a suitable brightness threshold is set to segment multiple levels of the retina. According to the grayscale features of the image, the boundaries of different retinal layers are extracted by edge detection technology. For example, the image can be classified at the pixel level by an automatic segmentation algorithm (such as a U-Net network based on deep learning), and the nerve fiber layer and inner core layer of the retina are extracted respectively. In the segmentation process, it is also possible to optimize in combination with the texture features of the local area to ensure the segmentation accuracy of each layer. Finally, the retinal layer data obtained includes nerve fiber layer data and inner core layer data, wherein the data of each layer includes its thickness, boundary information, and relative position, etc.

[0153] Step S43: performing thickness time series risk analysis on the retinal nerve fiber layer data, thereby obtaining nerve fiber layer thickness time risk data;

[0154] In this embodiment, the thickness data of the nerve fiber layer is processed in time series, and the thickness change of the nerve fiber layer at each time point is extracted. By comparing the data at multiple time points, the thickness change trend is analyzed. Statistical methods, such as linear regression or exponential smoothing methods, are used to calculate the rate of change of the thickness of the nerve fiber layer over time, and combined with a set risk threshold (such as a risk when the thickness decreases by more than a certain standard deviation value). Further, the data can be segmented and analyzed to identify time periods with more drastic thickness changes. Changes in the nerve fiber layer during these time periods indicate problems with the health of the retina. Through this analysis, the risk change interval of the retinal nerve fiber layer can be identified.

[0155] Step S44: performing regional thickness difference risk analysis on the retinal inner nuclear layer data, thereby obtaining regional thickness difference risk data of the inner nuclear layer;

[0156] In this embodiment, the inner nuclear layer of the retina is processed in blocks, and the entire retina is divided into multiple sub-regions. For each sub-region, its thickness is calculated and compared with the thickness of the surrounding area to obtain the thickness difference between the regions. A threshold is set, and when the thickness difference in a certain area exceeds the standard deviation range, the area is considered to be abnormal. In order to further accurately identify these areas, a method based on regional growth can be used to automatically divide areas of different thicknesses, and combined with local texture features, the significance of regional thickness differences is analyzed. Areas with large thickness differences indicate lesions in the inner nuclear layer, and further attention should be paid to changes in these areas.

[0157] Step S45: integrating the retinal thickness risk data according to the nerve fiber layer thickness time risk data and the inner nuclear layer regional thickness difference risk data, thereby obtaining the retinal thickness risk data.

[0158] In this embodiment, a weighted average or multidimensional data fusion algorithm is used to integrate the two types of data, and multidimensional factors such as thickness change rate, time difference, and regional difference are combined to obtain an overall retinal thickness risk index. During the integration process, different weight coefficients can be set according to the risk weights of different layers to ensure that the risks of the nerve fiber layer and the inner nuclear layer are reasonably considered. Further, risk classification is performed, for example, the retina is divided into low-risk, medium-risk, and high-risk areas based on the integrated data, so that doctors can promptly handle and monitor the retina according to the risk level. This comprehensive assessment method can more comprehensively and accurately assess the health status of the retina and identify potential disease risks in advance through multi-level data analysis.

[0159] Optionally, step S43 is specifically:

[0160] Step S431: Counting the thickness change of the retinal nerve fiber layer data, thereby obtaining the thickness change data of the retinal nerve fiber layer;

[0161] In this embodiment, the nerve fiber layer thickness value at each time point or each monitoring cycle is extracted from the obtained retinal nerve fiber layer data. The thickness data at each time point needs to be smoothed to eliminate noise interference. Common smoothing methods include median filtering and Gaussian filtering. Then, statistical analysis of the thickness data is performed at specified time intervals (such as monthly, quarterly). By calculating the change in thickness in each time period, the increase or decrease in the thickness of the nerve fiber layer is obtained, and a thickness change data set is generated. For example, the change trend of the thickness of the nerve fiber layer can be obtained by calculating the increment or difference in thickness of each cycle. This statistical process requires standardization of each data point to eliminate data deviations caused by equipment errors or different measurement conditions and ensure the reliability of the results.

[0162] Step S432: calculating the change rate according to the retinal nerve fiber layer thickness change data, thereby obtaining thickness change rate data;

[0163] In this embodiment, the thickness change rate is calculated based on the obtained retinal nerve fiber layer thickness change data. The calculation formula of the thickness change rate is:

[0164]

[0165] For each monitoring cycle, the ratio of thickness change between adjacent time periods is calculated. By calculating the rate of change at each time point, the trend of nerve fiber layer thickness change over time is further identified to obtain time series data. For example, if the thickness of the nerve fiber layer drops significantly within a certain cycle, the rate of change calculation will produce a large negative value, indicating a problem with retinal health. The data needs to be denoised, and the thickness change rate needs to be further smoothed using methods such as local regression or sliding average to avoid short-term abnormal fluctuations from interfering with the overall analysis results. This step provides a dynamic reflection of changes in nerve fiber layer health.

[0166] Step S433: obtaining thickness change rate risk threshold data;

[0167] In this embodiment, specific change rate thresholds are set according to the risk characteristics of different patient groups. These thresholds are usually determined by clinical experts based on a large amount of historical data or medical research results, reflecting the thickness variation range that the retinal nerve fiber layer should maintain under normal circumstances. For example, if the rate of change of the nerve fiber layer exceeds -10% or is less than +10%, it is considered to be at risk. This risk threshold can also be personalized according to different patient conditions, taking into account the impact of factors such as age and gender on retinal changes. The acquisition of threshold data usually comes from the guidance of medical literature or clinical practice, including different reference values ​​and standard deviation ranges. This data is the basis for subsequent risk assessment.

[0168] Step S434: performing risk classification on the thickness change rate data according to the thickness change rate risk threshold data, thereby obtaining nerve fiber layer thickness time risk data.

[0169] In this embodiment, the risk classification of the thickness change rate data is performed based on the set thickness change rate risk threshold data. The calculated thickness change rate data is compared with the set risk threshold, and the results are classified according to the set grading standards. For example, when the change rate exceeds -10%, it is considered that the change in the thickness of the nerve fiber layer exceeds the normal range and is in a high-risk state; when the change rate is between -10% and +10%, it is considered to be in the normal range and in a low-risk state; when the change rate exceeds +10%, it indicates that the lesion is in the early stage and the retinal health is in a low-risk state. Based on this division, the thickness time risk data of the nerve fiber layer is finally obtained, indicating the risk level at each time point or time period.

[0170] Optionally, step S44 is specifically:

[0171] Step S441: dividing the retinal inner nuclear layer data into regions, thereby obtaining the inner nuclear layer central region data and the inner nuclear layer peripheral region data;

[0172] In this embodiment, it is necessary to obtain a thickness distribution map of the inner nuclear layer of the retina. The central area and the peripheral area of ​​the inner nuclear layer are determined by segmenting the image and identifying the area. The specific operation steps include: first, the inner nuclear layer image of the retina is preliminarily processed using a threshold segmentation method to separate the inner nuclear layer of the retina from other layers; then, according to the spatial distribution of the inner nuclear layer, the central area and the peripheral area of ​​the inner nuclear layer are delineated according to certain standards. The central area is generally located in the center of the retina, while the peripheral area covers the part farther away from the central area. The boundary of the central area can be determined by calculating the geometric center of the image or using a clustering algorithm, and the peripheral area is defined by extending a certain distance from the outer edge of the central area. The standard for regional division needs to be set according to actual needs, such as using the pixel density or thickness value of the image to determine the boundary of the area.

[0173] Step S442: Calculating the average thickness according to the data of the central region of the inner core layer, thereby obtaining the average thickness data of the central region of the inner core layer;

[0174] In this embodiment, it is necessary to extract the thickness data of the central area, select all the pixels in the central area, and extract the thickness value corresponding to each pixel. Then, the arithmetic mean of all the thickness values ​​in the central area is calculated, and the formula is as follows:

[0175]

[0176] Among them, n is the number of pixels in the central area, and the thickness value i The thickness value of the i-th pixel. This calculation process ensures the overall evaluation of the thickness of the central area and eliminates the influence of local abnormal points on the results. When calculating, the data needs to be denoised, and the thickness value can be pre-processed using smoothing methods such as Gaussian filtering to improve the stability and accuracy of the results.

[0177] Step S443: Calculating the average thickness according to the data of the peripheral area of ​​the inner core layer, thereby obtaining the average thickness data of the peripheral area of ​​the inner core layer;

[0178] In this embodiment, the average thickness of the surrounding area is calculated based on the data of the surrounding area of ​​the kernel layer. The thickness data in the surrounding area is extracted. By defining the boundary of the area, the thickness values ​​of all pixels in the area are obtained, and the average thickness of the area is calculated. The calculation formula is:

[0179]

[0180] Among them, m is the number of pixels in the surrounding area, and the thickness value iis the thickness value of the i-th pixel in the surrounding area. In this process, the data in the surrounding area also needs to be preprocessed to remove noise and ensure the accuracy of the results. The quality of the data can be optimized through filtering, edge detection and other technologies to avoid the influence of extreme values ​​on the calculation results.

[0181] Step S444: performing thickness difference significance judgment based on the average thickness data of the central region of the inner core layer and the average thickness data of the peripheral region of the inner core layer, thereby obtaining thickness difference significance data;

[0182] In this embodiment, the thickness difference between the central area and the peripheral area is calculated:

[0183] Thickness difference = | average thickness of the central area - average thickness of the peripheral area |;

[0184] Then, use statistical methods, such as t-test or analysis of variance, to determine whether the thickness difference between the two is significant. Set the significance level (such as 0.05) and calculate the P value. If the P value is less than the significance level, the thickness difference between the central area and the peripheral area is considered to be significant, indicating that there are lesions or other abnormalities in the retina. The choice of statistical methods can be adjusted according to the distribution of the data and the sample size. In this process, the multiple comparison correction method of the hypothesis test can also be introduced as needed to avoid false positive results caused by too large a sample size.

[0185] Step S445: Perform risk assessment based on the thickness difference significance data to obtain inner core layer region thickness difference risk data.

[0186] In this embodiment, according to the result of significance judgment, patients with significant thickness difference are marked as high-risk groups. For patients with insignificant thickness difference, they are marked as low-risk groups. At this time, a risk scoring system can be set, for example, different risk levels are set according to the P value of the significance of the thickness difference, a P value less than 0.05 is high risk, a medium risk is a P value between 0.05 and 0.1, and a P value greater than 0.1 is low risk. In addition, according to clinical historical data and the specific circumstances of the patient, the risk assessment results can be further combined with other factors (such as age, gender, medical history, etc.) to make comprehensive adjustments to obtain the final inner nuclear layer regional thickness difference risk data. This risk data can provide doctors with a basis for decision-making and help identify patients with retinal diseases or abnormal changes.

[0187] Optionally, the present specification also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the computer program implements any one of the methods for training a vision risk prediction model based on multidimensional data.

[0188] Optionally, the present specification also provides a training system for a vision risk prediction model based on multidimensional data, which is used to execute the training method for a vision risk prediction model based on multidimensional data as described above. The training system for a vision risk prediction model based on multidimensional data includes:

[0189] Retinal vascular abnormality analysis module: used to obtain historical ophthalmological cases, and extract features of retinal fluorescence angiography images based on the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; perform retinal vascular abnormality analysis based on retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data;

[0190] Retinal abnormality recognition model construction module: used to construct a retinal abnormality recognition model based on retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; identify retinal abnormalities of patients in ophthalmic history cases based on the retinal abnormality recognition model, thereby obtaining retinal abnormality data of patients;

[0191] Fundus image risk assessment module: used to extract physiological test features of patients based on abnormal retinal data of patients, so as to obtain physiological test data of patients; perform fundus image risk assessment based on physiological test data of patients, so as to obtain fundus image risk data;

[0192] Retinal thickness risk assessment module: used to extract optical coherence tomography features based on the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image, and to perform retinal thickness risk assessment based on the optical coherence tomography image, thereby obtaining retinal thickness risk data;

[0193] Vision risk prediction model training module: used to train the vision risk prediction model based on retinal thickness risk data and fundus image risk data, so as to obtain a vision risk prediction training model.

[0194] The present invention provides a training system for a vision risk prediction model based on multidimensional data. The system can implement any training method for a vision risk prediction model based on multidimensional data of the present invention. The system is used to combine the operations between various modules and a medium for signal transmission to complete the training method for a vision risk prediction model based on multidimensional data. The internal modules of the system cooperate with each other to achieve more accurate vision risk prediction.

[0195] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0196] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A training method for a vision risk prediction model based on multidimensional data, characterized in that: The following steps are involved: Step S1: Obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; performing retinal vascular abnormality analysis according to the retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data; Step S2: constructing a retinal abnormality recognition model based on the retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; and performing retinal abnormality recognition on ophthalmic historical cases based on the retinal abnormality recognition model, thereby obtaining retinal abnormality data of the patient; Step S3: extracting physiological test features of the patient according to the abnormal retinal data of the patient, thereby obtaining physiological test data of the patient; performing fundus image risk assessment according to the physiological test data of the patient, thereby obtaining fundus image risk data; Step S4: extracting optical coherence tomography features according to the patient's retinal abnormality data to obtain an optical coherence tomography image, and performing retinal thickness risk assessment according to the optical coherence tomography image to obtain retinal thickness risk data; Step S5: Perform vision risk prediction model training based on the retinal thickness risk data and the fundus image risk data to obtain a vision risk prediction training model.

2. The method for training a vision risk prediction model based on multidimensional data according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: obtaining historical ophthalmological cases, and extracting features of retinal fluorescence angiography images according to the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; Step S12: performing vascular leakage analysis on the retinal fluorescence angiography image to obtain abnormal vascular leakage data; Step S13: performing vascular occlusion analysis on the retinal fluorescence angiography image to obtain abnormal vascular occlusion data; Step S14: performing retinal vascular abnormality integration according to the vascular leakage abnormality data and the vascular occlusion abnormality data, thereby obtaining retinal vascular abnormality data.

3. The training method of the vision risk prediction model based on multidimensional data according to claim 2, characterized in that: Step S12 is specifically as follows: Step S121: performing Gaussian filtering and smoothing on the retinal fluorescence angiography image, thereby obtaining a retinal fluorescence angiography smoothed image; Step S122: Calculate the horizontal gradient according to the retinal fluorescence angiography smoothing image, so as to obtain horizontal gradient data; Step S123: performing vertical gradient calculation according to the retinal fluorescence angiography smoothing image, thereby obtaining vertical gradient data; Step S124: performing gradient amplitude calculation according to the horizontal gradient data and the vertical gradient data to obtain gradient amplitude data, and performing gradient amplitude map drawing according to the gradient amplitude data to obtain a gradient amplitude map; Step S125: performing blood vessel edge detection according to the gradient amplitude map, thereby obtaining blood vessel edge data; Step S126: performing vascular leakage analysis on the vascular edge data to obtain vascular leakage abnormality data.

4. The method for training a vision risk prediction model based on multidimensional data according to claim 3, characterized in that: Step S126 is specifically as follows: Perform fluorescence intensity regional statistics on the blood vessel edge data to obtain high-intensity fluorescence area data; Performing a blood vessel expansion operation according to the high-intensity fluorescence region data, thereby obtaining blood vessel expansion data; Performing a blood vessel corrosion operation according to the blood vessel expansion data, thereby obtaining blood vessel corrosion data; Performing vascular expansion area identification on vascular corrosion data to obtain vascular expansion area data; Perform crack area identification on the vascular corrosion data to obtain vascular crack area data; The leakage anomaly is integrated according to the blood vessel expansion area data and the blood vessel crack area data, so as to obtain the blood vessel leakage anomaly data.

5. The method for training a vision risk prediction model based on multidimensional data according to claim 2, characterized in that: Step S13 is specifically as follows: Step S131: extracting retinal vascular area features from the retinal fluorescence angiography image, thereby obtaining retinal vascular area data; Step S132: calculating the blood vessel diameter according to the retinal blood vessel area data, thereby obtaining blood vessel diameter data; Step S133: marking the stenosis area of ​​the blood vessel diameter data, thereby obtaining the blood vessel stenosis area data; Step S134: Counting the fluorescence intensity of the vascular stenosis area data to obtain the fluorescence intensity data of the vascular stenosis area; Step S135: performing time change rate statistics based on the fluorescence intensity data of the vascular stenosis area, so as to obtain the vascular stenosis area with low change fluorescence intensity; Step S136: performing optical flow calculation based on the vascular stenosis area data, thereby obtaining the vascular stenosis area optical flow data; Step S137: extracting low optical flow vascular stenosis area features from the vascular stenosis area optical flow data, thereby obtaining low optical flow vascular stenosis area data; Step S138: performing a vascular occlusion anomaly intersection operation based on the low optical flow vascular stenosis area data and the low variation fluorescence intensity vascular stenosis area data, thereby obtaining vascular occlusion anomaly data.

6. The method for training a vision risk prediction model based on multidimensional data according to claim 1, characterized in that: Step S4 is specifically as follows: Step S41: extracting optical coherence tomography features according to the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image; Step S42: performing retinal layer segmentation on the optical coherence tomography image, thereby obtaining retinal nerve fiber layer data and retinal inner nuclear layer data; Step S43: performing thickness time series risk analysis on the retinal nerve fiber layer data, thereby obtaining nerve fiber layer thickness time risk data; Step S44: performing regional thickness difference risk analysis on the retinal inner nuclear layer data, thereby obtaining regional thickness difference risk data of the inner nuclear layer; Step S45: integrating the retinal thickness risk data according to the nerve fiber layer thickness time risk data and the inner nuclear layer regional thickness difference risk data, thereby obtaining the retinal thickness risk data.

7. The method for training a vision risk prediction model based on multidimensional data according to claim 6, characterized in that: Step S43 is specifically as follows: Step S431: Counting the thickness change of the retinal nerve fiber layer data, thereby obtaining the thickness change data of the retinal nerve fiber layer; Step S432: calculating the change rate according to the retinal nerve fiber layer thickness change data, thereby obtaining thickness change rate data; Step S433: obtaining thickness change rate risk threshold data; Step S434: performing risk classification on the thickness change rate data according to the thickness change rate risk threshold data, thereby obtaining nerve fiber layer thickness time risk data.

8. The method for training a vision risk prediction model based on multidimensional data according to claim 6, characterized in that: Step S44 is specifically as follows: Step S441: dividing the retinal inner nuclear layer data into regions, thereby obtaining the inner nuclear layer central region data and the inner nuclear layer peripheral region data; Step S442: Calculating the average thickness according to the data of the central region of the inner core layer, thereby obtaining the average thickness data of the central region of the inner core layer; Step S443: Calculating the average thickness according to the data of the peripheral area of ​​the inner core layer, thereby obtaining the average thickness data of the peripheral area of ​​the inner core layer; Step S444: performing thickness difference significance judgment based on the average thickness data of the central region of the inner core layer and the average thickness data of the peripheral region of the inner core layer, thereby obtaining thickness difference significance data; Step S445: Perform risk assessment based on the thickness difference significance data to obtain inner core layer region thickness difference risk data.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the training method of the vision risk prediction model based on multidimensional data as described in any one of claims 1 to 8 is implemented.

10. A training system for a vision risk prediction model based on multidimensional data, characterized in that: The method for training a vision risk prediction model based on multidimensional data according to claim 1, wherein the training system for the vision risk prediction model based on multidimensional data comprises: Retinal vascular abnormality analysis module: used to obtain historical ophthalmological cases, and extract features of retinal fluorescence angiography images based on the historical ophthalmological cases, thereby obtaining retinal fluorescence angiography images; perform retinal vascular abnormality analysis based on retinal fluorescence angiography images, thereby obtaining retinal vascular abnormality data; Retinal abnormality recognition model construction module: used to construct a retinal abnormality recognition model based on retinal vascular abnormality data, thereby obtaining a retinal abnormality recognition model; identify retinal abnormalities of patients in ophthalmic history cases based on the retinal abnormality recognition model, thereby obtaining retinal abnormality data of patients; Fundus image risk assessment module: used to extract physiological test features of patients based on abnormal retinal data of patients, so as to obtain physiological test data of patients; perform fundus image risk assessment based on physiological test data of patients, so as to obtain fundus image risk data; Retinal thickness risk assessment module: used to extract optical coherence tomography features based on the patient's retinal abnormality data, thereby obtaining an optical coherence tomography image, and to perform retinal thickness risk assessment based on the optical coherence tomography image, thereby obtaining retinal thickness risk data; Vision risk prediction model training module: used to train the vision risk prediction model based on retinal thickness risk data and fundus image risk data, so as to obtain a vision risk prediction training model.

Images