A method and apparatus for evaluating fundus lesions in chronic kidney disease

Abstract

The application discloses a fundus lesion evaluation method and device for chronic kidney disease, and the method comprises the following steps: acquiring an ultra-wide-angle fundus image of a target object to be measured with chronic kidney disease, and extracting fundus features from the ultra-wide-angle fundus image; based on the extracted fundus features, predicting a fundus lesion state by using a multivariate regression model; performing multimodal fusion processing on the fundus lesion state, patient information of the target object to be measured and clinical information; performing lesion probability prediction processing on the dynamic confidence interval of the multimodal fusion features to generate a chronic kidney disease fundus lesion risk grade. Therefore, not only can the chronic kidney disease fundus lesion risk grade be accurately predicted, but also the reliability of the chronic kidney disease fundus lesion risk grade evaluation is improved; moreover, an evaluation standard for the fundus lesion risk grade of early chronic kidney disease is constructed, so that the purpose of early detection and early treatment of patients is achieved, and irreversible fundus damage caused by CKD is reduced.

Description

Technical Field

[0001] This invention belongs to the field of fundus image processing technology, and particularly relates to a method and device for assessing fundus lesions in chronic kidney disease. Background Technology

[0002] Chronic kidney disease (CKD) often presents insidiously, with patients frequently experiencing no symptoms in the early stages, leading to low public awareness of the disease. However, as the disease progresses, patients may develop symptoms such as polyuria or fatigue due to anemia, marking a critical stage where the patient's risk increases significantly, with a substantial rise in the incidence of complications and the probability of progressing to end-stage renal disease. Diagnosis of CKD relies on biochemical indicators such as serum creatinine and eGFR, but these indicators have limitations: firstly, they are lagging indicators, often showing abnormalities only after significant kidney function impairment (usually in the stage of severe chronic renal failure); secondly, they are invasive, requiring blood tests and limiting frequent monitoring; and thirdly, there is a risk of missed diagnoses, as early renal microvascular lesions are difficult to detect using biochemical indicators. Furthermore, current technologies for assessing CKD primarily utilize renal imaging such as ultrasound (US), MRI, and CT scans, combined with radiomics or deep learning algorithms to evaluate CKD patients. For example, ultrasound texture features can be used to predict CKD staging (accuracy 75.95%), and MRI radiomics features can be used to predict ADPKD progression. The retina and other vital organs, such as the kidneys, share anatomical, embryological, and physiological connections. Researchers, leveraging the analytical capabilities of deep learning on retinal imaging, can explore the potential link between retinal microvascular changes and the early stages of chronic kidney disease. These findings are expected to reveal new prognostic biomarkers and risk stratification tools, helping clinicians intervene early to reduce the burden of chronic kidney disease-related complications. Previous studies have shown an association between elevated blood urea nitrogen and creatinine levels and the occurrence of specific eye diseases, including posterior subcapsular cataracts, advanced age-related macular degeneration, and diabetic retinopathy, suggesting a correlation between such ocular changes and altered renal function. Conversely, changes in fundus symptoms have also been found to potentially predict altered renal function.

[0003] Retinal lesions are a common complication in patients with chronic kidney disease (CKD), especially when CKD is caused by diabetes or hypertension, where the incidence is extremely high. Therefore, regular eye screenings are crucial for CKD patients. Currently, there is a lack of nationwide epidemiological studies of CKD and the establishment of automated predictive models. Early risk assessments of CKD retinal lesions largely rely on the subjective judgment of physicians' individual experience, which is prone to subjective bias. Because early retinal lesions in patients with mild CKD are rarely detected, objective and accurate prediction of the early risk of CKD-related retinal lesions is not possible.

[0004] Currently, CNN models built using OCTA can detect CKD. While this method achieves an AUC of 0.911, it has three limitations: first, detection is often limited to a single retinal layer, resulting in low accuracy and reliability; second, the examination cost is high; and third, OCTA requires a high degree of patient cooperation for ultra-wide-angle examinations. Because there is a lack of early non-invasive biomarkers for retinal complications in CKD patients, it is impossible to predict CKD in the subclinical stage. Furthermore, for stage 1-2 CKD patients, it is difficult to predict CKD combined with retinal disease using fundus images.

[0005] To address the aforementioned issues, there is an urgent need to develop a set of assessment criteria for the risk level of fundus lesions associated with early-stage chronic kidney disease. Summary of the Invention

[0006] This invention provides a method and device for assessing fundus lesions in chronic kidney disease; the method can accurately predict the risk of fundus lesions in early chronic kidney disease, thereby improving the accuracy of risk assessment of fundus lesions in chronic kidney disease.

[0007] According to a first aspect of the present invention, a method for assessing fundus lesions in patients with chronic kidney disease is provided. The method includes: acquiring an ultra-wide-angle fundus image of a target subject with chronic kidney disease, and extracting fundus features from the ultra-wide-angle fundus image; wherein the pre-selected fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease; predicting the fundus lesion status using a multivariate regression model based on the extracted fundus features; performing multimodal fusion processing based on the fundus lesion status, patient information of the target subject, and clinical information related to chronic kidney disease, and outputting multimodal fusion features; and performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level of fundus lesions in patients with chronic kidney disease.

[0008] Optionally, the step of predicting the state of fundus lesions using a multivariate regression model based on the extracted fundus features includes: sequentially performing data cleaning, outlier handling, feature standardization, and categorical variable transformation on the extracted fundus features to generate several preprocessed fundus features; calculating the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label; selecting preprocessed fundus features whose Pearson correlation coefficients meet preset conditions from the several preprocessed fundus features as candidate fundus features; performing LASSO regression dynamic screening on the candidate fundus features to generate quasi-fundus features; and using an elastic network model to predict the quasi-fundus features to generate the state of fundus lesions.

[0009] Optionally, the method further includes: obtaining quasi-fundus features corresponding to the ultra-wide-angle fundus image of the target object, and using the quasi-fundus features as training samples; wherein the target object is used to indicate the observed object with chronic kidney disease; training the model on several training samples based on an elastic network; adding L1+L2 double regularization terms during the model training process, and searching for the regularization strength through the network; adjusting the regularization strength by selecting 5-fold cross-validation, and generating an elastic network model.

[0010] Optionally, the step of performing multimodal fusion processing based on the fundus lesion state, the patient information of the target object, and the clinical information related to chronic kidney disease, and outputting multimodal fusion features includes: extracting clinical semantic features from the patient information of the target object and the clinical information related to chronic kidney disease, generating several candidate clinical semantic features; selecting candidate clinical semantic features associated with fundus lesions from the several candidate clinical semantic features as quasi-clinical features; and performing multimodal attention fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features.

[0011] Optionally, the step of extracting clinical semantic features from the patient information and chronic kidney disease-related clinical information of the target object to generate several candidate clinical semantic features includes: dividing different semantic groups according to clinical relevance based on the patient information and chronic kidney disease-related clinical information of the target object; the semantic groups include at least a renal function group, a microvascular injury group, and a disease course group; adding corresponding clinical semantic tags to each semantic group, and generating several candidate clinical semantic features based on at least one clinical semantic feature corresponding to each semantic group.

[0012] Optionally, the step of performing multimodal fusion processing based on the fundus lesion state, the patient information of the target object, and the clinical information related to chronic kidney disease to output multimodal fusion features includes: performing data cleaning, classification variable encoding, and data standardization processing on the patient information and the clinical information related to chronic kidney disease of the target object to generate standardized clinical data; calculating the covariance matrix of the standardized clinical data, and performing eigenvalue decomposition processing on the covariance matrix to obtain several clinical feature values ​​and a clinical feature vector corresponding to each clinical feature value; selecting clinical feature vectors whose clinical feature values ​​meet preset conditions from the several clinical feature vectors, and constructing a dimension reduction transformation matrix; performing dimension reduction mapping on the standardized clinical data based on the dimension reduction transformation matrix to generate quasi-clinical features; and performing multimodal fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features.

[0013] Optionally, the step of performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level for fundus lesions in chronic kidney disease includes: based on the multimodal fusion features, performing prediction processing using a gradient boosting tree to generate the probability of fundus lesions in chronic kidney disease; converting the measurement error corresponding to the quasi-clinical features and the measurement error corresponding to the fundus lesion state into the confidence level of the quasi-clinical features and the confidence level of the fundus lesion state, respectively; and determining the average confidence level corresponding to the multimodal fusion features based on the confidence level of the quasi-clinical features and the confidence level of the fundus lesion state; if the average confidence level is... If the number of specific features corresponding to a confidence level that satisfies the first preset condition and is less than a preset threshold satisfies the second preset condition, then a dynamic confidence interval corresponding to the multimodal fusion feature is generated. The dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval. The specific feature is used to indicate the feature selected from the quasi-clinical features and fundus lesion status used for multimodal fusion processing. Based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease, the dynamic probability interval is determined. Based on the dynamic probability interval, a risk level mapping is performed to generate a risk level for fundus lesions in chronic kidney disease.

[0014] Optionally, the step of mapping risk levels based on the dynamic probability interval to generate a risk level for chronic kidney disease fundus lesions includes: mapping the dynamic probability interval to risk levels based on preset rules to generate a level mapping result; if the level mapping result indicates that there is only one risk level mapped by the dynamic probability interval, then the risk level for chronic kidney disease fundus lesions is output; if the level mapping result indicates that there are two risk levels mapped by the dynamic probability interval, then the risk level for chronic kidney disease fundus lesions is determined based on quasi-clinical characteristics.

[0015] Optionally, the microvascular parameters include at least: arteriovenous diameter ratio, arterial fractal dimension, average arteriovenous tortuosity, average arteriovenous diameter, arteriovenous vascular density, fractal dimension of blood vessels in the subnasal region within a 1PD-3PD annular range, average tortuosity of blood vessels in the subnasal region within a 9mm circular range, and vascular density within a 13.5mm circular range.

[0016] The non-positive fundus indicators include at least: the distance from the macular center to the boundary, the coordinates of the continuous central line of the optic disc and macular region, the area and diameter of the optic disc, the area and diameter of the optic cup, the area of ​​the atrophic arc, and the density of the leopard spot. The positive fundus indicators include at least: neovascularization, microaneurysms, hemorrhage, hard osmosis, soft osmosis, drusen, epiretinal membrane, retinitis pigmentosa, retinal degeneration, retinal scleroderma, and fibrosis. The clinical information related to chronic kidney disease includes at least: blood pressure, renal biopsy results, urine albumin / creatinine ratio (UACR) (mg / g), red blood cells (HP), blood urea (BUN), blood creatinine (Scr), cystatin C, glomerular filtration rate (GFR), triglycerides (TG), total cholesterol (TC), hypertension, heart disease, and other chronic diseases and medication use.

[0017] According to a second aspect of the present invention, a device for assessing fundus lesions in chronic kidney disease is also provided. The device includes: an extraction module for acquiring an ultra-wide-angle fundus image of a target subject with chronic kidney disease and extracting fundus features from the ultra-wide-angle fundus image; wherein the pre-selected fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease; a first prediction module for predicting the state of fundus lesions based on the extracted fundus features using a multivariate regression model; a multimodal fusion module for performing multimodal fusion processing based on the fundus lesion state, patient information of the target subject, and clinical information related to chronic kidney disease, and outputting multimodal fusion features; and a second prediction module for performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level of fundus lesions in chronic kidney disease.

[0018] According to a third aspect of the present invention, an electronic device is also provided, comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the method as described in the first aspect.

[0019] According to a fourth aspect of the present invention, a computer-readable medium is also provided, on which a computer program is stored, wherein the program, when executed by a processor, implements the method described in the first aspect.

[0020] This invention provides a method and apparatus for assessing fundus lesions in patients with chronic kidney disease. The method includes: first, acquiring an ultra-wide-angle fundus image of a target subject with chronic kidney disease, and extracting fundus features from the ultra-wide-angle fundus image; wherein the pre-selected fundus features include positive fundus indicators, non-positive fundus indicators, and microvascular parameters related to chronic kidney disease; second, predicting the fundus lesion status using a multivariate regression model based on the extracted fundus features; then, performing multimodal fusion processing based on the fundus lesion status, patient information of the target subject, and clinical information related to chronic kidney disease, and outputting multimodal fusion features; finally, performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level of fundus lesions in patients with chronic kidney disease. This embodiment performs quantitative analysis of ultra-wide-angle fundus images of CKD patients to determine the state of fundus lesions. Then, it integrates the fundus lesion state, patient information, and clinical information of chronic kidney disease, and combines the lesion probability with dynamic confidence intervals to predict the risk level of fundus lesions in chronic kidney disease. As a result, it can accurately predict the risk level of fundus lesions in chronic kidney disease, improving the reliability of the risk level assessment of fundus lesions in chronic kidney disease. Moreover, it can move the diagnostic threshold of fundus lesions in CKD to earlier stages, which is conducive to early detection and early treatment of patients and reduces irreversible fundus damage caused by severe CKD. Attached Figure Description

[0021] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0022] Figure 1 This is a flowchart illustrating a method for assessing fundus lesions in chronic kidney disease, provided by an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of a process for generating a risk level of fundus lesions in chronic kidney disease according to an embodiment of the present invention;

[0024] Figure 3 A confidence mapping rule table provided for an embodiment of the present invention ( Figure 3 a) Interval Adjustment Rules Table ( Figure 3 b), and the probability interval and risk level mapping table ( Figure 3 c).

[0025] Figure 4 This is a schematic diagram of the structure of a fundus lesion assessment device for chronic kidney disease provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] Early-stage chronic kidney disease (CKD) can be difficult to detect. The method described in this embodiment aims not only to accurately predict the risk of CKD complications, but also to develop a set of assessment criteria for the risk level of CKD complications, thereby enabling patients to detect and treat the disease early.

[0028] like Figure 1 The diagram shown is a flowchart illustrating a method for assessing fundus lesions in chronic kidney disease according to an embodiment of the present invention.

[0029] A method for assessing fundus lesions in chronic kidney disease includes at least the following steps:

[0030] S101, acquire ultra-wide-angle fundus images of the target subject with chronic kidney disease, and extract fundus features from the ultra-wide-angle fundus images; wherein, the pre-selected fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease;

[0031] S102, Based on the extracted fundus features, a multivariate regression model is used to predict the state of fundus lesions;

[0032] S103, based on the state of fundus lesions, patient information of the target subject, and clinical information related to chronic kidney disease, performs multimodal fusion processing and outputs multimodal fusion features;

[0033] S104 performs dynamic confidence interval prediction on the multimodal fusion features to generate the risk level of fundus lesions in chronic kidney disease.

[0034] In step S101, ultra-wide-angle fundus images of both eyes of the target subject are acquired using a non-mydriatic fundus camera; the ultra-wide-angle fundus images are refractively corrected to obtain corrected fundus images; motion artifact elimination processing is performed on the corrected fundus images based on a reference fundus image of the target subject to generate quasi-fundus images; wherein, the reference fundus image is used to indicate a clear and artifact-free ultra-wide-angle fundus image corresponding to the same eye position of the target subject; arteriovenous segmentation processing is performed on the quasi-fundus images to generate fundus vessel segmentation results; microvascular parameters related to chronic kidney disease are extracted from the fundus vessel segmentation results; and fundus positive and non-positive indicators are extracted from the quasi-fundus images.

[0035] Microvascular parameters are used to indicate the quantitative characteristics of microvessels, including structural parameters (e.g., morphology) and functional parameters (e.g., blood flow). Non-positive fundus indicators indicate normal fundus structures or physiological features and have no direct pathological significance. Positive fundus indicators indicate pathological changes in the fundus and directly suggest lesions.

[0036] For example, the microvascular parameters include at least: arteriovenous diameter ratio, arterial fractal dimension, average arteriovenous tortuosity, average arteriovenous diameter, arteriovenous vascular density, fractal dimension of blood vessels in the subnasal region within a 1PD-3PD annular range, average tortuosity of blood vessels in the subnasal region within a 9mm circular range, and vascular density within a 13.5mm circular range.

[0037] The non-positive indicators of the fundus include at least: the distance from the center of the macula to the boundary, the coordinates of the continuous central line of the optic disc and macula, the area and diameter of the optic disc, the area and diameter of the optic cup, the area of ​​the atrophic arc, and the density of the leopard spot;

[0038] The positive indicators of the fundus include at least: neovascularization, microaneurysm, hemorrhage, hard osmosis, soft osmosis, drusen, epiretinal membrane, retinal pigment changes, retinal degeneration, cataracts, and fibrosis.

[0039] Here, ultra-wide-angle fundus images serve as the basic data for examining fundus lesions. Not only is data acquisition convenient, but multiple microvascular parameters related to chronic kidney disease can also be obtained.

[0040] In S102, positive and non-positive fundus indicators, as well as microvascular parameters related to chronic kidney disease, are input into a multivariate regression model to output the fundus lesion status. The fundus lesion status can be represented by HR values ​​or by a scoring system; however, no specific limitations are imposed here.

[0041] For example, the step of predicting the state of fundus lesions using a multivariate regression model based on the extracted fundus features includes: sequentially performing data cleaning, outlier handling, feature standardization, and categorical variable transformation on the extracted fundus features to generate several preprocessed fundus features; calculating the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label; selecting preprocessed fundus features whose Pearson correlation coefficients meet preset conditions from the several preprocessed fundus features as candidate fundus features; performing LASSO regression dynamic screening on the candidate fundus features to generate quasi-fundus features; and using an elastic network model to predict the quasi-fundus features to generate the state of fundus lesions. Here, the elastic network model is pre-trained.

[0042] For example: Data cleaning: For missing features (such as "vascular density within a 13.5mm circular area" in microvascular parameters), the "grouped mean imputation method" is used: the mean of this feature is calculated separately for each CKD stage (stages 1-5), and the missing values ​​in the same group are filled with the mean of the corresponding stage (reducing the error by more than 30% compared to global mean imputation).

[0043] Outlier handling: If the ratio of arterial to venous diameter is greater than 1:1 (physiologically, the diameter of an artery should be smaller than that of a vein), it is judged as a measurement error and automatically corrected to the 95th percentile value of the same batch of samples (e.g., 0.7).

[0044] Feature standardization: The Z-score standardization method is used for standardization.

[0045] Categorical variable transformation: For binary variables such as "history of hypertension" (yes / no), one-hot coding is used to convert them into 0 / 1 values.

[0046] The fundus lesion label, as the dependent variable, includes: a label indicating the presence of fundus lesions (dependent variable = 1) and a label indicating the absence of fundus lesions (dependent variable = 0). The Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label is calculated. From several preprocessed fundus features, those with a Pearson correlation coefficient greater than a preset threshold are selected as candidate fundus features. Therefore, this filtering operation can retain fundus features strongly correlated with fundus lesions, thereby reducing the computational cost of the elastic network model and avoiding interference from weakly correlated features in model learning and prediction.

[0047] First, LASSO regression was used to calculate feature coefficients. Second, 5-fold cross-validation was used to determine the optimal feature coefficients. Then, based on the sparsity property of L1 regularization, candidate fundus features with zero feature coefficients were forcibly deleted, while candidate fundus features with non-zero feature coefficients were retained to generate quasi-fundus features. Thus, dynamic screening through LASSO regression can further eliminate redundant features; for example, "arterial classification dimension" is highly correlated with "arterial vascular density," so only the latter is retained.

[0048] As further exemplarily, the elastic network model is obtained through the following method: acquiring the quasi-fundus features corresponding to the ultra-wide-angle fundus image of the target object, and using the quasi-fundus features as training samples; wherein, the target object is used to indicate the observed object with chronic kidney disease; training the model on several training samples based on the elastic network; adding L1+L2 double regularization terms during the model training process, and searching for the regularization strength through the network; adjusting the regularization strength by selecting 5-fold cross-validation, and generating the elastic network model.

[0049] This embodiment further enhances feature selection through L1 regularization, thereby ensuring that only core quasi-fundus features participate in the prediction; and reduces the impact of collinearity of at least two feature coefficients through L2 regularization, thereby improving the reliability of fundus lesion state prediction.

[0050] In S103, for example, the step of performing multimodal fusion processing based on the fundus lesion state, the patient information of the target object, and the clinical information related to chronic kidney disease, and outputting multimodal fusion features includes: extracting clinical semantic features from the patient information of the target object and the clinical information related to chronic kidney disease, generating several candidate clinical semantic features; selecting candidate clinical semantic features associated with fundus lesions from the several candidate clinical semantic features as quasi-clinical features; and performing multimodal attention fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features.

[0051] As further exemplified, the step of extracting clinical semantic features from the patient information and chronic kidney disease-related clinical information of the target object to generate several candidate clinical semantic features includes: dividing different semantic groups according to clinical relevance based on the patient information and chronic kidney disease-related clinical information of the target object; the semantic groups include at least a renal function group, a microvascular injury group, and a disease course group; adding corresponding clinical semantic tags to each semantic group, and generating several candidate clinical semantic features based on at least one clinical semantic feature corresponding to each semantic group.

[0052] Specifically, the clinical information related to chronic kidney disease includes at least: blood pressure, renal biopsy results, urine albumin / creatinine ratio (UACR) (mg / g), red blood cells (HP), blood urea nitrogen (BUN), blood creatinine (Scr), cystatin C, glomerular filtration rate (GFR), triglycerides (TG), total cholesterol (TC), hypertension, heart disease and other chronic diseases, and medication use.

[0053] The patient information includes: age, CKD course, height, and weight.

[0054] For example: Renal function group:

[0055] 1. GFR: <30mL / min / 1.73m 2 → "Renal failure", 30-59 mL / min / 1.73 m 2 → "Renal insufficiency", 60-90 mL / min / 1.73 m 2 → "Mildly decreased renal function", >90 mL / min / 1.73 m 2 →“Normal kidney function”;

[0056] 2. Serum creatinine: For men, 0.6-1.2 mg / dL (converted to normal range of 53-106 μg / L) → "Normal creatinine"; >1.2 mg / dL → "Elevated creatinine";

[0057] For women, creatinine levels between 0.5-1.1 mg / dL (converted to a normal range of 44-97 umol / L) indicate "normal creatinine"; levels >1.1 mg / dL indicate "elevated creatinine".

[0058] Microvascular injury group:

[0059] 1. UACR: >300mg / g → “severe proteinuria”, 30-300mg / g → “microalbuminuria”, <30mg / g → “negative proteinuria”;

[0060] 2. Systolic blood pressure: >140mmHg → "Hypertension", ≤140mmHg → "Normal blood pressure".

[0061] Disease course group:

[0062] 1. Age: >60 years old → "elderly", 40-60 years old → "middle-aged", <40 years old → "youth";

[0063] 2. CKD course: >10 years → "long course", 3-10 years → "medium course", <3 years → "short course".

[0064] When the patient information and chronic kidney disease-related clinical information of the target subject are "GFR=25, serum creatinine=1, UACR=400, age=65, disease duration=12, systolic blood pressure=150", the labeling is "renal failure, severe proteinuria, elderly, long disease duration, hypertension".

[0065] The same semantic group retains features that are more strongly associated with fundus lesions.

[0066] For example: in the renal function group, GFR was better than serum creatinine, so GFR was retained and serum creatinine was removed;

[0067] Microvascular injury group: UACR and systolic blood pressure were preserved;

[0068] Disease course group: Age and CKD disease course were retained.

[0069] Finally, the five quasi-clinical characteristics were output as follows: GFR=25 (renal failure), UACR=400 (severe proteinuria), systolic blood pressure=150 (hypertension), age=65 (elderly), and disease duration=12 (long disease duration).

[0070] This embodiment overcomes the limitations of traditional data dimensionality reduction and black box models; it realizes the interpretable flow from raw data to risk level, solves the problem of lack of medical logic in pure data-driven methods, and improves the reliability of fundus lesion assessment in chronic kidney disease.

[0071] For example, the step of performing multimodal fusion processing based on the fundus lesion state, patient information of the target subject, and clinical information related to chronic kidney disease to output multimodal fusion features includes: performing data cleaning, categorical variable encoding, and data standardization on the patient information and clinical information related to chronic kidney disease of the target subject to generate standardized clinical data; calculating the covariance matrix of the standardized clinical data and performing eigenvalue decomposition on the covariance matrix to obtain several clinical feature values ​​and a clinical feature vector corresponding to each clinical feature value; selecting clinical feature vectors whose clinical feature values ​​meet preset conditions from the several clinical feature vectors and constructing a dimension reduction transformation matrix; performing dimension reduction mapping on the standardized clinical data based on the dimension reduction transformation matrix to generate quasi-clinical features; and performing multimodal fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features. This overcomes the influence of data structure, effectively fuses the fundus lesion state, patient information of the target subject, and clinical information related to chronic kidney disease, and improves the reliability of fundus lesion assessment in chronic kidney disease.

[0072] In S104, based on preset rules or models, the multimodal fusion features are subjected to dynamic confidence interval prediction of lesion probability to generate the risk level of fundus lesions in chronic kidney disease.

[0073] The method in this embodiment quantifies CKD-specific microvascular damage by combining retinal microvascular parameters, which not only enables accurate screening of high-risk groups, but also solves the problem of multimodal data fusion and improves the diagnostic accuracy of fundus lesions in patients with comorbidities.

[0074] The effects of this embodiment will be explained from the following aspects:

[0075] The method in this embodiment uses quantitative analysis of ultra-wide-angle fundus images of CKD patients to determine fundus lesions. This not only moves the diagnostic threshold for fundus lesions in CKD forward, but also enables early warning.

[0076] The method described in this embodiment not only avoids repeated blood draws and is suitable for dynamic monitoring, but is also more convenient than US / MRI / CT examinations, can generate reports in real time, and is applicable to primary healthcare, guiding primary healthcare personnel in prefecture-level cities to assess the fundus condition of CKD patients; thus, it enables non-invasive and convenient assessment of fundus lesions in chronic kidney disease.

[0077] The method described in this embodiment can significantly improve the accuracy of predicting the state of fundus lesions in CKD patients, thereby achieving precise stratified prediction.

[0078] The method described in this embodiment can enable early detection and intervention of fundus lesions in chronic kidney disease, thereby reducing severe visual impairment caused by fundus lesions in patients with moderate to severe CKD, preventing blindness, and thus reducing the treatment costs for patients.

[0079] like Figure 2 The diagram shown is a flowchart illustrating the process of generating risk levels for fundus lesions in chronic kidney disease according to an embodiment of the present invention.

[0080] The multimodal fusion features are processed with dynamic confidence intervals to predict the probability of lesions, generating a risk level for fundus lesions in chronic kidney disease; including:

[0081] S201, based on multimodal fusion features, uses gradient boosting trees for prediction processing to generate the probability of fundus lesions in chronic kidney disease;

[0082] S202, the measurement errors corresponding to quasi-clinical features and fundus lesion states are converted into confidence scores for quasi-clinical features and fundus lesion states, respectively; and based on the confidence scores for quasi-clinical features and fundus lesion states, the average confidence score corresponding to the multimodal fusion features is determined.

[0083] S203, if the number of specific features corresponding to the average confidence level that meets the first preset condition and is less than the preset threshold meets the second preset condition, then a dynamic confidence interval corresponding to the multimodal fusion feature is generated; the dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval; the specific feature is used to indicate the features selected from the quasi-clinical features and fundus lesion status used for multimodal fusion processing;

[0084] S204, Based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease, determine the dynamic probability interval;

[0085] S205 generates risk levels for retinopathy in chronic kidney disease by mapping risk levels based on dynamic probability intervals.

[0086] In S205, for example, risk level mapping is performed based on dynamic probability intervals to generate risk levels for retinopathy of chronic kidney disease; including: mapping risk levels on dynamic probability intervals based on preset rules to generate level mapping results; if the level mapping results indicate that there is only one risk level mapped by the dynamic probability interval, then the risk level for retinopathy of chronic kidney disease is output; if the level mapping results indicate that there are two risk levels mapped by the dynamic probability interval, then the risk level for retinopathy of chronic kidney disease is determined based on quasi-clinical characteristics.

[0087] For example: Acquire several features for generating multimodal fusion features, such as quasi-clinical features and fundus lesion status. Extract the nominal error of each feature from the detection device log (e.g., GFR nominal error ±5% from the biochemical analyzer, UACR nominal error ±15% from the urine protein detector); calculate the actual fluctuation error (the difference rate between two test values ​​within 3 months for the same patient) based on historical test data. If the actual error > nominal error, the actual error is used (e.g., if a patient's actual UACR fluctuation is 20%, then 20% is used). Then, based on... Figure 3 The measurement error value is converted into a confidence level of 0-1.0 (the smaller the error, the higher the confidence level). Therefore, through "equipment traceability + historical verification," the error is dynamically updated, improving accuracy by 35% compared to traditional fixed error values ​​(such as using only equipment specifications).

[0088] Calculate the confidence scores of quasi-clinical features and the mean confidence scores of fundus lesion states; count the number of features with low confidence scores. Then, based on the mean confidence score and the number of features with low confidence scores, compare... Figure 3 b. Determine the dynamic confidence interval corresponding to the multimodal fusion features. Substitute the probability of fundus lesions in chronic kidney disease into the dynamic confidence interval to determine the dynamic probability interval. For example: if the average confidence level is >0.8 and there is only one low-confidence feature (UACR=0.7), then the dynamic probability interval is the 95% confidence interval [P-0.05, P+0.05].

[0089] Comparison Figure 3 c. Map the dynamic probability interval to risk levels to generate a level mapping result. If the level mapping result indicates that there is only one risk level mapped by the dynamic probability interval, then the risk level of chronic kidney disease fundus lesions based on the risk level mapping is output; if the level mapping result indicates that there are two risk levels mapped by the dynamic probability interval, then the risk level of chronic kidney disease fundus lesions is determined based on quasi-clinical features. For example: when the dynamic probability interval spans multiple levels, quasi-clinical feature verification is initiated. For example: P = 0.58, the dynamic probability interval [0.53, 0.63] simultaneously covers levels 3 and 4; if GFR < 30 or fundus neovascularization exists, then the level is increased by 1 level; if UACR < 30 and blood pressure is normal, then the level is decreased by 1 level; if there are no special indicators, then the level output by the level mapping table is maintained, that is, 1. Based on the mapping between the core probability interval and the risk level, the risk level of chronic kidney disease fundus lesions is output as level 3. Thus, the accuracy of the risk level assessment of chronic kidney disease fundus lesions is improved by the dual-track mapping mechanism of "probability interval master rule + clinical feature verification rule".

[0090] This embodiment achieves a precise and interpretable mapping from multimodal fusion features to risk levels by combining dynamic confidence intervals with clinical rules. It is particularly suitable for scenarios such as retinopathy in chronic kidney disease, where a balance between data quantification and clinical experience is required. The significant effects are as follows:

[0091] 1. Accuracy of grading: Multicenter data validation shows that the consistency of this method with ophthalmologist diagnosis reaches 91%, which is 12% higher than the traditional fixed probability grading (79%).

[0092] 2. Clinical suitability: Dynamic intervals ensure that the grading is not affected by low-quality data (e.g., UACR with large errors will widen the interval to avoid over-grading);

[0093] 3. Interpretability: Each level includes probabilistic justification, data quality, and clinical feature contribution, which aligns with physicians' decision-making habits.

[0094] It should be noted that, Figure 3 a, Figure 3 b, and Figure 3 c is derived from medical research, expert clinical experience, a large amount of clinical case data, and a comprehensive consideration of data characteristics, aiming to balance the accuracy of risk prediction with clinical applicability.

[0095] The following is a detailed description of a method for assessing fundus lesions in chronic kidney disease provided in this embodiment, using specific application scenarios.

[0096] A method for assessing fundus lesions in chronic kidney disease includes at least the following steps:

[0097] S1, acquire ultra-wide-angle fundus images of the target subject with chronic kidney disease, and extract fundus features from the ultra-wide-angle fundus images; wherein, the pre-selected fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease;

[0098] S2, the extracted fundus features are sequentially cleaned, outlier-handled, feature-standardized, and categorical variable-transformed to generate several preprocessed fundus features; the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label is calculated; preprocessed fundus features whose Pearson correlation coefficients meet preset conditions are selected as candidate fundus features from the several preprocessed fundus features; the candidate fundus features are dynamically screened using LASSO regression to generate quasi-fundus features; the quasi-fundus features are predicted using an elastic network model to generate the fundus lesion status.

[0099] S3. Based on the patient information and chronic kidney disease-related clinical information of the target object, different semantic groups are divided according to clinical relevance; the semantic groups include at least the renal function group, the microvascular injury group, and the disease course group; corresponding clinical semantic labels are added to each semantic group, and several candidate clinical semantic features are generated based on at least one clinical semantic feature corresponding to each semantic group; candidate clinical semantic features associated with fundus lesions are selected from the several candidate clinical semantic features as quasi-clinical features; multimodal attention fusion processing is performed on the quasi-clinical features and the fundus lesion status to generate multimodal fusion features.

[0100] S4, based on multimodal fusion features, uses gradient boosting trees for prediction processing to generate the probability of fundus lesions in chronic kidney disease.

[0101] S5, the measurement errors corresponding to the quasi-clinical features and the fundus lesion states are converted into confidence levels of the quasi-clinical features and fundus lesion states, respectively; and based on the confidence levels of the quasi-clinical features and the fundus lesion states, the average confidence level corresponding to the multimodal fusion features is determined; if the number of specific features corresponding to the average confidence level that meets the first preset condition and is less than the preset threshold meets the second preset condition, then a dynamic confidence interval corresponding to the multimodal fusion features is generated; the dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval; the specific features are used to indicate the features selected from the quasi-clinical features and fundus lesion states used for multimodal fusion processing.

[0102] S6. Based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease, determine the dynamic probability interval; map the dynamic probability interval to risk levels based on preset rules to generate a level mapping result; if the level mapping result indicates that the dynamic probability interval maps to only one risk level, then output the risk level of fundus lesions in chronic kidney disease; if the level mapping result indicates that the dynamic probability interval maps to two risk levels, then determine the risk level of fundus lesions in chronic kidney disease based on quasi-clinical characteristics.

[0103] Compared to predicting the risk level of chronic kidney disease (CKD) retinopathy solely based on the state of fundus lesions, the method in this embodiment significantly improves the sensitivity of CKD retinopathy risk prediction. Compared to predicting CKD retinopathy risk based on simplified combinations of microvascular parameters (e.g., using only arteriovenous diameter ratio, average venous tortuosity, venous density, etc.), this method is more universal and can cover various medical scenarios with varying medical resources. Compared to predicting CKD retinopathy risk solely by dynamically monitoring the annual decline rate of GFR, this method can shift the diagnostic focus of CKD-associated retinopathy to earlier stages, enabling early warning.

[0104] Therefore, this embodiment achieves interpretable mapping of "fundus lesion status - clinical indicators - risk level" by deeply embedding medical association algorithms. This not only solves the problem of subjective bias caused by the fact that early CKD fundus lesion risk assessment mostly relies on doctors' personal experience, but also accurately predicts the risk level of early CKD fundus lesions.

[0105] like Figure 4 The diagram shown is a structural schematic of a fundus lesion assessment device for chronic kidney disease provided in an embodiment of the present invention.

[0106] A device for assessing fundus lesions in chronic kidney disease, the device 400 comprising: an extraction module 401 for acquiring ultra-wide-angle fundus images of a target subject with chronic kidney disease and extracting fundus features from the ultra-wide-angle fundus images; wherein the pre-selected fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease; a first prediction module 402 for predicting the state of fundus lesions based on the extracted fundus features using a multivariate regression model; a multimodal fusion module 403 for performing multimodal fusion processing based on the fundus lesion state, patient information of the target subject, and clinical information related to chronic kidney disease, and outputting multimodal fusion features; and a second prediction module 404 for performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level of fundus lesions in chronic kidney disease.

[0107] In a preferred embodiment of this example, the first prediction module includes: a transformation processing unit, used to sequentially perform data cleaning, outlier processing, feature standardization, and categorical variable transformation on the extracted fundus features to generate several preprocessed fundus features; a calculation unit, used to calculate the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label; a selection unit, used to select preprocessed fundus features whose Pearson correlation coefficient meets preset conditions from the several preprocessed fundus features as candidate fundus features; a screening unit, used to perform LASSO regression dynamic screening on the candidate fundus features to generate quasi-fundus features; and a prediction processing unit, used to use an elastic network model to predict the quasi-fundus features to generate the fundus lesion state.

[0108] In a preferred embodiment of this invention, the device further includes: an acquisition unit, configured to acquire quasi-fundus features corresponding to an ultra-wide-angle fundus image of the target object, and use the quasi-fundus features as training samples; wherein the target object is used to indicate an observed object with chronic kidney disease; a model training unit, configured to train a model based on an elastic network using several training samples; a search unit, configured to add an L1+L2 double regularization term during model training and search for the regularization strength through the network; and a generation unit, configured to select 5-fold cross-validation to adjust the regularization strength and generate an elastic network model.

[0109] In a preferred embodiment of this example, the multimodal fusion module includes: an extraction unit, used to extract clinical semantic features from patient information and clinical information related to chronic kidney disease of the target object to be tested, and generate a number of candidate clinical semantic features; a selection unit, used to select candidate clinical semantic features related to fundus lesions from the number of candidate clinical semantic features as quasi-clinical features; and a generation unit, used to perform multimodal attention fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features.

[0110] In a preferred embodiment of this example, the extraction unit includes: a segmentation subunit, used to segment different semantic groups according to clinical relevance based on the patient information of the target object and the clinical information related to chronic kidney disease; the semantic groups include at least a renal function group, a microvascular injury group, and a disease course group; and a generation subunit, used to add corresponding clinical semantic tags to each semantic group and generate several candidate clinical semantic features based on at least one clinical semantic feature corresponding to each semantic group.

[0111] In a preferred embodiment of this example, the multimodal fusion module includes: a preprocessing unit, used to perform data cleaning, classification variable encoding, and data standardization on patient information and chronic kidney disease-related clinical information of the target object to generate standardized clinical data; a generation unit, used to calculate the covariance matrix of the standardized clinical data and perform eigenvalue decomposition on the covariance matrix to obtain several clinical feature values ​​and a clinical feature vector corresponding to each clinical feature value; a selection unit, used to select clinical feature vectors whose clinical feature values ​​meet preset conditions from the several clinical feature vectors and construct a dimension reduction transformation matrix; a dimension reduction mapping unit, used to perform dimension reduction mapping on the standardized clinical data based on the dimension reduction transformation matrix to generate quasi-clinical features; and a multimodal fusion processing unit, used to perform multimodal fusion processing on the quasi-clinical features and the fundus lesion state to generate multimodal fusion features.

[0112] In a preferred embodiment of this example, the second prediction module includes: a prediction processing unit, configured to perform prediction processing using a gradient boosting tree based on the multimodal fusion features to generate the probability of fundus lesions in chronic kidney disease; a first determination unit, configured to convert the measurement error corresponding to the quasi-clinical features and the measurement error corresponding to the fundus lesion state into the confidence scores of the quasi-clinical features and the fundus lesion state, respectively; and to determine the average confidence score corresponding to the multimodal fusion features based on the confidence scores of the quasi-clinical features and the fundus lesion state; and a first generation unit, configured to generate a prediction module if the average confidence score satisfies a first preset condition. If the number of specific features corresponding to a confidence level less than a preset threshold satisfies a second preset condition, then a dynamic confidence interval corresponding to the multimodal fusion feature is generated; the dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval; the specific feature is used to indicate the feature selected from the quasi-clinical features and fundus lesion status used for multimodal fusion processing; a second determining unit is used to determine the dynamic probability interval based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease; a second generating unit is used to perform risk level mapping based on the dynamic probability interval to generate a risk level for fundus lesions in chronic kidney disease.

[0113] In a preferred embodiment of this example, the second generation unit includes: a generation subunit, used to map the dynamic probability interval to risk levels based on preset rules and generate a level mapping result; an output subunit, used to output the risk level of chronic kidney disease fundus lesions if the level mapping result indicates that there is only one risk level mapped by the dynamic probability interval; and a determination subunit, used to determine the risk level of chronic kidney disease fundus lesions based on quasi-clinical characteristics if the level mapping result indicates that there are two risk levels mapped by the dynamic probability interval.

[0114] In a preferred embodiment of this example, the microvascular parameters include at least: arteriovenous diameter ratio, arterial fractal dimension, average arteriovenous tortuosity, average arteriovenous diameter, arteriovenous vascular density, fractal dimension of blood vessels in the subnasal region within a 1PD-3PD annular range, average tortuosity of blood vessels in the subnasal region within a 9mm circular range, and vascular density within a 13.5mm circular range.

[0115] The non-positive indicators of the fundus include at least: the distance from the center of the macula to the boundary, the coordinates of the continuous central line of the optic disc and macula, the area and diameter of the optic disc, the area and diameter of the optic cup, the area of ​​the atrophic arc, and the density of the leopard spot;

[0116] The positive indicators of the fundus include at least: neovascularization, microaneurysms, hemorrhage, hard osmosis, soft osmosis, drusen, epiretinal membrane, retinal pigment changes, retinal degeneration, cataracts, and fibrosis.

[0117] The clinical information related to chronic kidney disease includes at least: blood pressure, renal biopsy results, urine albumin / creatinine ratio (UACR) (mg / g), red blood cells (HP), blood urea nitrogen (BUN), blood creatinine (Scr), cystatin C, glomerular filtration rate (GFR), triglycerides (TG), total cholesterol (TC), hypertension, heart disease and other chronic diseases, and medication use.

[0118] The aforementioned device for assessing fundus lesions in chronic kidney disease can execute a method for assessing fundus lesions in chronic kidney disease provided in an embodiment of the present invention, and possesses the corresponding functional modules and beneficial effects for executing such a method. Technical details not described in detail in this embodiment can be found in the method for assessing fundus lesions in chronic kidney disease provided in an embodiment of the present invention.

[0119] The present invention also provides an electronic device, comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the present invention's method for assessing fundus lesions in chronic kidney disease.

[0120] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.

[0121] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0122] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the methods according to the following embodiments of this application described in the "Exemplary Methods" section above.

[0123] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0124] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0125] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0126] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0127] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0128] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

[0129] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0130] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0131] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assessing fundus lesions in chronic kidney disease, characterized in that, include: Acquire ultra-wide-angle fundus images of target subjects with chronic kidney disease, and extract fundus features from the ultra-wide-angle fundus images; wherein, the fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease; Based on the extracted fundus features, a multivariate regression model is used to predict the state of fundus lesions; Based on the fundus lesion status, the patient information of the target subject, and the clinical information related to chronic kidney disease, multimodal fusion processing is performed to output multimodal fusion features; The multimodal fusion features are subjected to dynamic confidence interval prediction of lesion probability to generate a risk level of fundus lesions in chronic kidney disease; The method for predicting fundus lesion status based on extracted fundus features using a multivariate regression model includes: sequentially performing data cleaning, outlier handling, feature standardization, and categorical variable transformation on the extracted fundus features to generate several preprocessed fundus features; calculating the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label; selecting preprocessed fundus features whose Pearson correlation coefficients meet preset conditions from the several preprocessed fundus features as candidate fundus features; performing LASSO regression dynamic screening on the candidate fundus features to generate quasi-fundus features; and using an elastic network model to predict the quasi-fundus features to generate the fundus lesion status. The step of performing dynamic confidence interval prediction processing on the multimodal fusion features to generate a risk level for fundus lesions in chronic kidney disease includes: based on the multimodal fusion features, performing prediction processing using a gradient boosting tree to generate the probability of fundus lesions in chronic kidney disease; converting the measurement error corresponding to the quasi-clinical feature and the measurement error corresponding to the fundus lesion state into the confidence level of the quasi-clinical feature and the confidence level of the fundus lesion state, respectively; and determining the average confidence level corresponding to the multimodal fusion feature based on the confidence level of the quasi-clinical feature and the confidence level of the fundus lesion state; if the average confidence level meets a first preset condition and is less than a preset threshold confidence level... If the number of specific features meets the second preset condition, a dynamic confidence interval corresponding to the multimodal fusion feature is generated; the dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval; the specific feature is used to indicate the feature selected from the quasi-clinical features and fundus lesion status used for multimodal fusion processing; based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease, a dynamic probability interval is determined; based on the dynamic probability interval, a risk level mapping is performed to generate a risk level for fundus lesions in chronic kidney disease; wherein, the quasi-clinical features are determined based on the patient information of the target subject and the clinical information related to chronic kidney disease.

2. The method according to claim 1, characterized in that, Also includes: The quasi-fundus features corresponding to the ultra-wide-angle fundus image of the target object are obtained, and the quasi-fundus features are used as training samples; wherein, the target object is used to indicate the observed object with chronic kidney disease; Based on elastic networks, model training is performed on several training samples; During model training, L1+L2 double regularization terms are added, and the regularization strength is searched through the network. We selected 5-fold cross-validation to adjust the regularization strength and generated an elastic network model.

3. The method according to claim 1, characterized in that, Based on the fundus lesion status, patient information of the target subject, and clinical information related to chronic kidney disease, multimodal fusion processing is performed to output multimodal fusion features; including: Clinical semantic features are extracted from patient information and chronic kidney disease-related clinical information of the target subjects to generate several candidate clinical semantic features. From the aforementioned candidate clinical semantic features, candidate clinical semantic features associated with fundus lesions are selected as quasi-clinical features; The quasi-clinical features and the fundus lesion state are subjected to multimodal attention fusion processing to generate multimodal fusion features.

4. The method according to claim 3, characterized in that, The process involves extracting clinical semantic features from patient information and chronic kidney disease-related clinical information of the target subject, generating several candidate clinical semantic features, including: Based on the patient information and chronic kidney disease-related clinical information of the target subjects, different semantic groups are divided according to clinical relevance; the semantic groups include at least the renal function group, the microvascular injury group, and the disease course group. Add corresponding clinical semantic tags to each semantic group, and generate several candidate clinical semantic features based on at least one clinical semantic feature corresponding to each semantic group.

5. The method according to claim 1, characterized in that, Based on the fundus lesion status, patient information of the target subject, and clinical information related to chronic kidney disease, multimodal fusion processing is performed to output multimodal fusion features; including: The patient information and chronic kidney disease-related clinical information of the target subjects are cleaned, categorical variable coding is performed, and data standardization is carried out to generate standardized clinical data. Calculate the covariance matrix of standardized clinical data, and perform eigenvalue decomposition on the covariance matrix to obtain several clinical feature values ​​and a clinical feature vector corresponding to each clinical feature value; From the aforementioned clinical feature vectors, select clinical feature vectors whose clinical feature values ​​meet preset conditions, and construct a dimension reduction transformation matrix; Based on the dimensionality reduction transformation matrix, the standardized clinical data are dimensionality-reduced and mapped to generate quasi-clinical features; The quasi-clinical features and the fundus lesion status are subjected to multimodal fusion processing to generate multimodal fusion features.

6. The method according to claim 1, characterized in that, The process of mapping risk levels based on the dynamic probability interval to generate risk levels for retinopathy in chronic kidney disease includes: Based on preset rules, the dynamic probability interval is mapped to a risk level to generate a level mapping result; If the level mapping result represents only one risk level of the dynamic probability interval mapping, then the risk level of chronic kidney disease fundus lesions is output. If the grade mapping result represents two risk levels of the dynamic probability interval mapping, then the risk level of the chronic kidney disease fundus lesions is determined based on quasi-clinical characteristics.

7. The method according to claim 1, characterized in that, The microvascular parameters include at least: arteriovenous diameter ratio, arterial fractal dimension, average arteriovenous tortuosity, average arteriovenous diameter, arteriovenous vascular density, fractal dimension of blood vessels in the subnasal region within a 1PD-3PD annular range, average tortuosity of blood vessels in the subnasal region within a 9mm circular range, and vascular density within a 13.5mm circular range. The non-positive indicators of the fundus include at least: the distance from the center of the macula to the boundary, the coordinates of the continuous central line of the optic disc and macula, the area and diameter of the optic disc, the area and diameter of the optic cup, the area of ​​the atrophic arc, and the density of the leopard spot; The positive indicators of the fundus include at least: neovascularization, microaneurysms, hemorrhage, hard osmosis, soft osmosis, drusen, epiretinal membrane, retinal pigment changes, retinal degeneration, cataracts, and fibrosis. The clinical information related to chronic kidney disease includes at least: blood pressure, renal biopsy results, urine albumin / creatinine ratio (UACR) mg / g, red blood cells / HP, blood urea / BUN, blood creatinine / Scr, cystatin C, glomerular filtration rate / GFR, triglycerides / TG, total cholesterol / TC, hypertension, heart disease and other chronic diseases, and medication use.

8. A device for assessing fundus lesions in patients with chronic kidney disease, characterized in that, include: An extraction module is used to acquire ultra-wide-angle fundus images of target subjects with chronic kidney disease and extract fundus features from the ultra-wide-angle fundus images; wherein, the fundus features include fundus positive indicators, fundus non-positive indicators, and microvascular parameters related to chronic kidney disease; The first prediction module is used to predict the state of fundus lesions based on the extracted fundus features using a multivariate regression model. The multimodal fusion module is used to perform multimodal fusion processing based on the fundus lesion status, patient information of the target object, and clinical information related to chronic kidney disease, and output multimodal fusion features. The second prediction module is used to perform dynamic confidence interval prediction processing on the multimodal fusion features to generate the risk level of fundus lesions in chronic kidney disease. The first prediction module includes: a transformation processing unit, used to sequentially perform data cleaning, outlier handling, feature standardization, and categorical variable transformation on the extracted fundus features to generate several preprocessed fundus features; a calculation unit, used to calculate the Pearson correlation coefficient between each preprocessed fundus feature and the fundus lesion label; a selection unit, used to select preprocessed fundus features whose Pearson correlation coefficients meet preset conditions from the several preprocessed fundus features as candidate fundus features; a screening unit, used to perform LASSO regression dynamic screening on the candidate fundus features to generate quasi-fundus features; and a prediction processing unit, used to use an elastic network model to predict the quasi-fundus features to generate the fundus lesion status. The second prediction module includes: a prediction processing unit, used to perform prediction processing based on the multimodal fusion features using a gradient boosting tree to generate the probability of fundus lesions in chronic kidney disease; a first determination unit, used to convert the measurement error corresponding to the quasi-clinical features and the measurement error corresponding to the fundus lesion state into the confidence scores of the quasi-clinical features and the fundus lesion state, respectively; and to determine the average confidence score corresponding to the multimodal fusion features based on the confidence scores of the quasi-clinical features and the fundus lesion state; and a first generation unit, used to generate a number of specific features corresponding to the confidence scores that satisfy a first preset condition and are less than a preset threshold, which satisfy a second preset condition. Given the conditions, a dynamic confidence interval corresponding to the multimodal fusion feature is generated; the dynamic confidence interval is used to indicate the dynamic range corresponding to the 95% confidence interval; the specific feature is used to indicate the feature selected from the quasi-clinical features and fundus lesion status used for multimodal fusion processing; wherein, the quasi-clinical features are determined based on the patient information of the target object and the clinical information related to chronic kidney disease; the second determining unit is used to determine the dynamic probability interval based on the dynamic confidence interval and the probability of fundus lesions in chronic kidney disease; the second generating unit is used to perform risk level mapping based on the dynamic probability interval to generate the risk level of fundus lesions in chronic kidney disease.

Images