A chorioidal-fundus image-based brain choroid plexus volume prediction method and system

Abstract

This invention discloses a method and system for predicting the volume of the choroid plexus of the brain based on fundus images. The method includes a model building stage and a target object prediction stage: In the model building stage, the brain magnetic resonance imaging (MRI) image, fundus image, and basic information of the training object are acquired; the choroid plexus region is segmented based on the MRI image and the volume label of the choroid plexus is determined; the fundus image is processed to generate input data for the training model, and a choroid plexus volume prediction model is constructed; In the target object prediction stage, the fundus image and basic information of the target object are acquired; the model input data is generated and input into the choroid plexus volume prediction model to obtain the predicted value of the choroid plexus volume, the probability of high choroid plexus volume overload, or the risk level of high choroid plexus volume overload. This invention can achieve auxiliary prediction of the state of the choroid plexus volume based on fundus images without directly inputting MRI images.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and artificial intelligence-assisted assessment technology, specifically to a method and system for predicting the volume of the choroid plexus of the brain based on fundus images. Background Technology

[0002] The choroid plexus is an important structure located within the ventricular system, mainly composed of blood vessels and epithelial cells. It participates in cerebrospinal fluid production, substance transport, metabolic waste removal, and the maintenance of cerebral microenvironment homeostasis. The volume of the choroid plexus can reflect its structural state and is associated with cerebrospinal fluid dynamics, changes in the cerebral microenvironment, and various brain imaging phenotypes. Previous studies have shown that the volume of the choroid plexus is associated with neurodegenerative changes, cognitive decline, abnormal cerebrospinal fluid circulation, and altered brain clearance function (see Jiang J, Zhuo Z, Wang A, et al. Choroid plexus volume as a novel candidate neuroimaging marker of the Alzheimer's continuum[J].Alzheimer's Research & Therapy, 2024; Lin W, Chen H, Zhang Z, et al.Multimodal MRI reveals impaired glymphatic function with choroid plexus enlargement and cerebrospinal fluid expansion in Alzheimer's disease[J].Scientific Reports, 2025).

[0003] Currently, the volume of the choroid plexus is typically obtained using magnetic resonance imaging (MRI). For example, the choroid plexus region within the ventricles can be segmented based on T1-weighted images, and the volume of the choroid plexus can be determined according to the segmented region. However, MRI examinations are costly, time-consuming, and require specialized equipment and personnel, making them unsuitable as a routine assessment method for large-scale population screening or in primary healthcare settings. For individuals who have not yet undergone MRI examinations, relying solely on MRI images to obtain the volume of the choroid plexus would increase the burden on healthcare resources.

[0004] Fundus imaging can non-invasively, rapidly, and cost-effectively display retinal vessels and fundus microcirculation structures. As an extension of the central nervous system, the retina has certain connections with brain tissue in terms of vascular regulation, microcirculation, and fluid circulation. Previous studies have attempted to use retinal image analysis to predict brain imaging abnormalities such as white matter hyperintensities (see Lau AY, Mok V, Lee J, et al. Retinal image analytics detects white matter hyperintensities in healthy adults[J]. Annals of Clinical and Translational Neurology, 2019), suggesting a modelable relationship between fundus images and brain imaging phenotypes, providing a technical basis for using fundus images to conduct brain imaging feature-assisted screening.

[0005] However, existing technologies have not yet established a complete scheme for predicting the volume of the choroid plexus based on fundus images, and mainly have the following shortcomings: First, existing brain imaging feature prediction methods based on fundus images mainly focus on cerebrovascular injury-related phenotypes such as white matter hyperintensity, and have not yet taken the volume of the choroid plexus as the main prediction object; Second, existing studies on the volume of the choroid plexus mainly rely on the segmentation and calculation of brain magnetic resonance images, making it difficult to obtain choroid plexus volume prediction information when the target subject has not undergone brain magnetic resonance examination. Summary of the Invention

[0006] To address the lack of a complete scheme for predicting the volume of the choroid plexus based on fundus images in existing technologies, this invention provides a method and system for predicting the volume of the choroid plexus based on fundus images. This method determines the volume of the choroid plexus of the training subject using magnetic resonance imaging (MRI) images, and constructs a choroid plexus volume label or a high-overload label based on the choroid plexus volume. It obtains at least one of the following through fundus image processing: preprocessed fundus images, fundus image processing results, quantitative features of fundus images, and basic information. Furthermore, it constructs a choroid plexus volume prediction model based on the mapping relationship between the fundus image data of the training subject and the choroid plexus volume label, thereby achieving auxiliary prediction of the choroid plexus volume state of the target subject.

[0007] This invention provides a method for predicting the volume of the choroid plexus in the brain based on fundus images, comprising the following steps:

[0008] 1) Obtain fundus images of the target subject;

[0009] 2) The fundus images are processed to obtain model input data for predicting the volume of the choroid plexus in the brain; wherein the model input data includes at least one of the following: preprocessed fundus images, fundus image processing results, quantitative features of fundus images, and basic information;

[0010] 3) Input the model input data into the pre-constructed brain choroid plexus volume prediction model to obtain the brain choroid plexus volume prediction information of the target object;

[0011] The brain choroid plexus volume prediction model is constructed based on the mapping relationship between the fundus image data of the training subjects and the brain choroid plexus volume label, wherein the brain choroid plexus volume label is determined by the brain choroid plexus region in the brain magnetic resonance image of the training subjects.

[0012] The brain choroid plexus volume prediction information includes at least one of the following: brain choroid plexus volume prediction value, brain choroid plexus volume high overload probability, and brain choroid plexus volume high overload risk level.

[0013] The fundus images include at least one of fundus photographic images, optical coherence tomography (OCT) angiography images, and OCT structural images.

[0014] Processing the fundus image includes at least one of the following steps: performing image quality control on the fundus image; performing at least one preprocessing on the fundus image, including image standardization, brightness correction, contrast enhancement, noise suppression, and artifact correction; and locating, segmenting, or extracting at least one of the fundus structural region, vascular region, blood flow signal region, and retinal layer structural region according to the type of the fundus image.

[0015] In one optional embodiment, when the fundus image is a fundus photographic image, processing the fundus image includes at least one of optic disc region localization, macular region localization, retinal vessel region segmentation, vessel enhancement, vessel binarization, vessel skeletonization, and arteriovenous classification; when the fundus image is an optical coherence tomography (OCT) vascular imaging image, processing the fundus image includes at least one of blood flow signal extraction, vessel region segmentation, retinal or choroidal vessel layer segmentation, perfusion region extraction, and avascular region extraction; when the fundus image is an OCT structural image, processing the fundus image includes at least one of retinal layer segmentation, choroidal region segmentation, macular region localization, and optic disc region localization.

[0016] The model input data includes at least one of the following: preprocessed fundus image, fundus image processing result, quantitative features of fundus image, and basic information; wherein, the fundus image processing result includes at least one of the following corresponding to the fundus image type: retinal vessel segmentation map, retinal artery segmentation map, retinal vein segmentation map, vascular skeleton map, blood flow signal map, vascular layer segmentation map, perfusion area map, avascular area map, retinal layer segmentation map, and choroidal region segmentation map.

[0017] The quantitative features of the fundus image include at least one of the following: vascular morphology features, blood perfusion features, layer structure features, and regional morphology features corresponding to the fundus image type.

[0018] The quantitative features of the fundus image corresponding to the fundus photograph include at least one of the following: fractal dimension, vascular density, vascular width, average vascular width, vascular tortuosity, vascular branching characteristics, and arteriovenous width ratio of at least one vascular region among the whole blood vessels, arteries, and veins.

[0019] The quantitative features of the fundus image corresponding to the optical coherence tomography angiography image include at least one of the following: retinal vessel density, perfusion density, blood flow area, avascular area, blood flow defect area, blood flow defect ratio, superficial vascular plexus vessel density, deep vascular plexus vessel density, peridiscal capillary density, and choroidal capillary layer blood flow defect parameters.

[0020] The quantitative features of the fundus image corresponding to the optical coherence tomography structural image include at least one of the following: retinal thickness, choroidal thickness, retinal nerve fiber layer thickness, ganglion cell complex thickness, macular thickness, macular volume, and optic disc-related structural parameters.

[0021] In one alternative implementation, the basic information includes at least one of age, gender, years of education, body mass index, blood pressure status, diabetes status, smoking status, and alcohol consumption status.

[0022] The process of determining the brain choroid plexus volume label includes: acquiring brain magnetic resonance images of the training subjects; segmenting the brain choroid plexus region from the brain magnetic resonance images; determining the brain choroid plexus volume based on the brain choroid plexus region, and obtaining the brain choroid plexus volume label.

[0023] The brain magnetic resonance image includes a T1-weighted image; the brain choroid plexus volume label is the brain choroid plexus volume determined based on the brain choroid plexus region segmented from the brain magnetic resonance image.

[0024] The probability of high choroid plexus volume overload or the risk level of high choroid plexus volume overload is determined based on a high choroid plexus volume overload label, which is generated based on the relationship between the choroid plexus volume of the training subject and a preset threshold. The preset threshold includes at least one of a quantile threshold of choroid plexus volume in the training set and a choroid plexus volume threshold determined based on a normal reference population. The risk level of high choroid plexus volume overload includes at least two of low risk, medium risk, and high risk.

[0025] The process of constructing the brain choroid plexus volume prediction model includes: acquiring fundus image data and brain magnetic resonance images of multiple training subjects; determining the brain choroid plexus volume based on the brain magnetic resonance images of the training subjects; determining the training model input data based on the fundus image data of the training subjects; using the training model input data as input, and using at least one of the brain choroid plexus volume, brain choroid plexus volume high overload label, and brain choroid plexus volume high overload risk level label as supervision information, to construct the brain choroid plexus volume prediction model.

[0026] In one optional embodiment, the brain plexus volume prediction model includes at least one of a machine learning model, a deep learning model, and a fusion model; the machine learning model includes at least one of support vector machine, logistic regression, random forest, gradient boosting decision tree, LightGBM, XGBoost, CatBoost, multilayer perceptron, ridge regression, LASSO regression, and elastic network regression; the deep learning model includes at least one of convolutional neural network, residual network, DenseNet, EfficientNet, visual Transformer, Swing Transformer, U-Net encoder network, attention neural network, and multimodal fusion neural network.

[0027] The present invention also provides a brain choroid plexus volume prediction system based on fundus images, comprising:

[0028] The image acquisition module is used to acquire fundus images of the target object;

[0029] The image processing module is used to process the fundus images to obtain model input data for predicting the volume of the choroid plexus in the brain.

[0030] The prediction module is used to input the model input data into a pre-constructed brain choroid plexus volume prediction model to obtain brain choroid plexus volume prediction information of the target object;

[0031] The result output module is used to output the predicted brain choroid plexus volume information.

[0032] The present invention also provides an electronic device and a computer-readable storage medium. The electronic device includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned method for predicting the volume of the choroid plexus based on fundus images. The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the above-mentioned method for predicting the volume of the choroid plexus based on fundus images.

[0033] The beneficial effects of this invention are:

[0034] 1) This invention uses the volume of the choroid plexus determined by brain magnetic resonance imaging as supervision information, so that the model output corresponds to the brain imaging structural index of the choroid plexus volume, thereby improving the radiological interpretability of the prediction results.

[0035] 2) This invention extracts fundus image features corresponding to the image type from fundus images and can combine them with basic information to construct model input data, enabling the target object to obtain brain choroid plexus volume prediction information without directly inputting brain magnetic resonance images, thus improving the convenience of auxiliary assessment of brain choroid plexus volume status.

[0036] 3) The model input data of the present invention may include at least one of the following: preprocessed fundus images, fundus image processing results, quantitative features of fundus images, and basic information, and can be compatible with different types of fundus images, artificial features, image features, and structured information.

[0037] 4) The brain choroid plexus volume prediction model of the present invention can be a machine learning model, a deep learning model or a fusion model. It can be modeled using interpretable quantitative features of fundus images or end-to-end deep learning or multimodal fusion modeling, and has a wide range of applications.

[0038] 5) The output results of this invention may include at least one of the following: predicted choroid plexus volume, probability of high choroid plexus volume overload, and risk level of high choroid plexus volume overload. These results can be used to assist in the assessment of the choroid plexus volume status and provide a reference for whether to conduct further MRI examination, follow-up assessment, or manual review. Attached Figure Description

[0039] Figure 1 This is an overall flowchart of the present invention, used to illustrate the main steps of the model building stage S101 to S106 and the target object prediction stage S201 to S203.

[0040] Figure 2 A schematic diagram illustrating the construction of volume labels for the choroid plexus in the brain, wherein, Figure 2 In the image (a), the image is a T1-weighted image. Figure 2(b) in the image shows the segmented choroid plexus, used to illustrate the process of segmenting the choroid plexus region based on T1-weighted images and determining the volume of the choroid plexus.

[0041] Figure 3 This is a schematic diagram of the brain choroid plexus volume prediction model, used to show the input, feature processing, fusion, and output relationships of machine learning models, deep learning models, and fusion models.

[0042] Figure 4 The graph shows the prediction performance of the choroid plexus volume prediction model, which is used to display the prediction performance of the choroid plexus volume prediction model for the quartile states of the choroid plexus volume. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0044] In this embodiment of the invention, the "target object" can be a subject who needs to undergo auxiliary assessment of the choroid plexus volume status; the "training object" can be a sample object that has both fundus imaging data and brain magnetic resonance imaging data; the "choroid plexus volume prediction information" is not used as a disease diagnosis conclusion, but is used to provide auxiliary assessment information on the choroid plexus volume status.

[0045] Example 1: Construction of a model for predicting brain choroid plexus volume

[0046] This embodiment provides a method for constructing a brain choroid plexus volume prediction model based on fundus images, such as... Figure 1 As shown, it includes the following steps.

[0047] S101, Obtain sample data from multiple training objects.

[0048] The sample data includes fundus imaging data, brain magnetic resonance imaging (MRI) images, and basic information of the training subjects. The fundus imaging data includes at least one of fundus photographs, optical coherence tomography (OCT) angiography images, and OCT structural images. The brain MRI images include T1-weighted images. The basic information includes at least one of age, sex, years of education, body mass index, blood pressure status, diabetes status, smoking status, and alcohol consumption status.

[0049] In one specific implementation, the fundus image data is a fundus photograph, and the basic information includes age, gender, and years of education. The fundus photograph can be a central image of the optic disc, a central image of the macula, or an image that simultaneously includes the optic disc and macula.

[0050] S102, Process the brain magnetic resonance imaging image to determine the volume of the choroid plexus.

[0051] Specifically, the T1-weighted image of the training object is obtained; the choroid plexus region is segmented from the T1-weighted image; and the volume of the choroid plexus is determined according to a preset volume calculation rule based on the choroid plexus region.

[0052] like Figure 2 As shown, Figure 2 In the image (a), the image is a T1-weighted image. Figure 2 (b) in the image shows the segmented choroid plexus. By segmenting the choroid plexus region on the T1-weighted image, the choroid plexus region used to determine the volume of the choroid plexus can be obtained. The choroid plexus region can be obtained through manual segmentation, semi-automatic segmentation, or automatic segmentation methods. Automatic segmentation methods can employ threshold-based, atlas-based, morphological processing-based, machine learning-based, or deep learning-based segmentation methods.

[0053] S103, construct brain choroid plexus volume labels based on brain choroid plexus volume.

[0054] When the model is used to output continuous numerical values, the choroid plexus volume is used as the choroid plexus volume label. When the model is used to output the probability of high choroid plexus volume overload or the risk level of high choroid plexus volume overload, a high choroid plexus volume overload label is generated based on the relationship between the training subject's choroid plexus volume and a preset threshold.

[0055] The preset threshold includes at least one of the quantile thresholds of the brain choroid plexus volume in the training set and the brain choroid plexus volume threshold determined based on a normal reference population. In one specific embodiment, samples whose brain choroid plexus volume is located in the upper quartile of the training set are marked as high-load samples of brain choroid plexus volume, and the remaining samples are marked as low-load samples.

[0056] S104, process the fundus image data of the training subjects to obtain the input data for the training model.

[0057] Specifically, image quality control and preprocessing are performed on the fundus image data. Image quality control includes determining whether the fundus images exhibit significant blurring, underexposure, overexposure, visual field defects, occlusion, motion artifacts, or layering errors. Fundus images that do not meet quality requirements can be discarded, re-acquired, or flagged with quality warnings. Preprocessing includes at least one of image normalization, brightness correction, contrast enhancement, noise suppression, and artifact correction.

[0058] Subsequently, based on the type of fundus image data, at least one of the fundus structural region, vascular region, blood flow signal region, and retinal layer structural region is located, segmented, or extracted to obtain fundus image processing results corresponding to the fundus image type.

[0059] In one specific implementation, when the fundus image data is a fundus photograph image, the processing includes at least one of optic disc region localization, macular region localization, retinal vessel region segmentation, vessel enhancement, vessel binarization, vessel skeletonization, and arteriovenous classification, and can obtain at least one of retinal vessel segmentation map, retinal artery segmentation map, retinal vein segmentation map, and vessel skeleton map.

[0060] In another specific embodiment, when the fundus image data is an optical coherence tomography angiography image, the processing includes at least one of blood flow signal extraction, vascular region segmentation, retinal or choroidal vascular layer segmentation, perfusion region extraction and avascular region extraction, and at least one of blood flow signal map, vascular layer segmentation map, perfusion region map and avascular region map can be obtained.

[0061] In another specific embodiment, when the fundus image data is an optical coherence tomography (OCT) structural image, the processing includes at least one of retinal layer segmentation, choroidal region segmentation, macular region localization, and optic disc region localization, and at least one of the retinal layer segmentation map and choroidal region segmentation map can be obtained.

[0062] Furthermore, quantitative features of the fundus images are extracted based on the fundus image processing results. These quantitative features include at least one of the following: vascular morphology features, blood flow perfusion features, layer structure features, and regional morphology features corresponding to the fundus image type.

[0063] When the fundus image data is a fundus photograph, the quantitative features of the fundus image may include at least one of the following: fractal dimension, vessel density, vessel width, average vessel width, vessel tortuosity, vessel branching characteristics, and arteriovenous width ratio of at least one vascular region among the whole vessels, arteries, and veins. Preferably, the quantitative features of the fundus image include fractal dimension, vessel density, arterial fractal dimension, arterial vessel density, and average vein width, denoted as FD, VD, AFD, AVD, and VAW, respectively.

[0064] When the fundus image data is an optical coherence tomography (OCT) vascular imaging image, the quantitative features of the fundus image may include at least one of the following: retinal vessel density, perfusion density, blood flow area, avascular area, blood flow defect area, blood flow defect ratio, superficial vascular plexus vessel density, deep vascular plexus vessel density, peridiscal capillary density, and choroidal capillary layer blood flow defect parameters.

[0065] When the fundus image data is an optical coherence tomography (OCT) structural image, the quantitative features of the fundus image may include at least one of the following: retinal thickness, choroidal thickness, retinal nerve fiber layer thickness, ganglion cell complex thickness, macular thickness, macular volume, and optic disc-related structural parameters.

[0066] S105, Generate input data for training the model.

[0067] At least one of the following can be used as input data for training the model: preprocessed fundus image, fundus image processing result, quantitative features of fundus image, and basic information.

[0068] In one specific implementation, the input data for training the model includes FD, VD, AFD, AVD, VAW, as well as age, gender, and years of education.

[0069] S106, perform model training to obtain a brain choroid plexus volume prediction model.

[0070] Using the training model input data as input, and at least one of the following as supervision information—brain choroid plexus volume, brain choroid plexus volume high-overload label, and brain choroid plexus volume high-overload risk level label—the brain choroid plexus volume prediction model is trained.

[0071] When the supervised information is the volume of the choroid plexus, the choroid plexus volume prediction model can be constructed as a regression model to output the predicted value of the choroid plexus volume. When the supervised information is a label indicating high choroid plexus volume overload, the choroid plexus volume prediction model can be constructed as a classification model to output the probability of high choroid plexus volume overload. When the supervised information is a label indicating the risk level of high choroid plexus volume overload, the choroid plexus volume prediction model can be constructed as a multi-class classification model or an ordered classification model to output the risk level of high choroid plexus volume overload.

[0072] The brain plexus volume prediction model includes at least one of machine learning models, deep learning models, and fusion models. The machine learning model includes at least one of support vector machines, logistic regression, random forests, gradient boosting decision trees, LightGBM, XGBoost, CatBoost, multilayer perceptrons, ridge regression, LASSO regression, and elastic network regression. The deep learning model includes at least one of convolutional neural networks, residual networks, DenseNet, EfficientNet, visual Transformer, Swin Transformer, U-Net encoder network, attention neural networks, and multimodal fusion neural networks.

[0073] The trained choroid plexus volume prediction model can be deployed on local computers, medical imaging workstations, servers, cloud platforms, or mobile terminals for predicting the choroid plexus volume of target subjects.

[0074] Example 2: Prediction of choroid plexus volume in target brain

[0075] This embodiment provides a method for predicting the volume of the choroid plexus in the brain based on fundus images, such as... Figure 1 As shown, it includes the following steps.

[0076] S201, Obtain target object data.

[0077] The target object data includes fundus images and basic information of the target object. The fundus images include at least one of fundus photography images, optical coherence tomography (OCT) angiography images, and OCT structural images. The basic information includes at least one of age, gender, years of education, body mass index, blood pressure status, diabetes status, smoking status, and alcohol consumption status.

[0078] In one specific implementation, the fundus image of the target object is a monocular or binocular fundus photograph. When acquiring binocular fundus images, the fundus images of the left and right eyes can be processed separately, and the prediction results of both eyes can be averaged, maximized, weighted and fused, or output separately.

[0079] S202, Generate model input data.

[0080] Based on the processing method determined in the model construction phase, the fundus images of the target object are processed to generate model input data. Specifically, at least one of the following processes is performed on the fundus images of the target object: quality control, image preprocessing, structural region localization, vascular region segmentation, blood flow signal extraction, or retinal layer structure segmentation, and quantitative features of the fundus images corresponding to the fundus image type are extracted.

[0081] The model input data includes at least one of the following: preprocessed fundus images, fundus image processing results, quantitative features of fundus images, and basic information.

[0082] In one specific implementation, the model input data includes FD, VD, AFD, AVD, VAW, as well as age, gender, and years of education.

[0083] S203, input the prediction model and get the prediction results.

[0084] Input the model input data into the trained brain choroid plexus volume prediction model to obtain the brain choroid plexus volume prediction information of the target object.

[0085] The brain choroid plexus volume prediction information includes at least one of the following: brain choroid plexus volume prediction value, brain choroid plexus volume high overload probability, and brain choroid plexus volume high overload risk level.

[0086] In one specific implementation, the choroid plexus volume prediction model outputs a high-risk probability of choroid plexus volume. When the high-risk probability is less than a first threshold, a low-risk probability is output; when the high-risk probability is greater than or equal to the first threshold and less than a second threshold, a medium-risk probability is output; and when the high-risk probability is greater than or equal to the second threshold, a high-risk probability is output. The first and second thresholds can be determined based on the training set, validation set, or application scenario.

[0087] The predicted choroid plexus volume information can be used as an auxiliary assessment result of the choroid plexus volume status, and should not be used as a disease diagnosis conclusion. For target subjects whose prediction results indicate a high risk of high choroid plexus volume overload, it may be suggested to combine clinical data or further brain MRI examination for a comprehensive judgment.

[0088] Example 3: Structure of a brain choroid plexus volume prediction model

[0089] This embodiment provides a brain choroid plexus volume prediction model structure, such as Figure 3 As shown.

[0090] The model input data includes at least one of the following: preprocessed fundus images or fundus image processing results, quantitative features of fundus images, and basic information. Preprocessed fundus images or fundus image processing results can be input into a deep learning model, while quantitative features of fundus images and basic information can be input into a machine learning model or into the corresponding branch of a fusion model.

[0091] In one implementation, preprocessed fundus images or fundus image processing results are input into a deep learning model. The deep learning model may include at least one of convolutional neural networks, residual networks, Transformers, or attention networks. The deep learning model is used to automatically extract deep features from fundus images and output brain choroid plexus volume prediction information.

[0092] In another implementation, quantitative features and basic information from fundus images are input into a machine learning model. The machine learning model may include at least one of support vector machines, random forests, XGBoost, and LightGBM. The machine learning model is used to output predicted choroid plexus volume information based on structured input data.

[0093] In another implementation, the fusion model includes an image branch, a quantitative feature branch, and a basic information branch. The image branch processes preprocessed fundus images or fundus image processing results; the quantitative feature branch processes quantitative features of fundus images; and the basic information branch processes basic information such as age, gender, and years of education. The features output from each branch are concatenated, weighted, or attention-based and then input into the choroid plexus volume prediction model to obtain choroid plexus volume prediction information.

[0094] like Figure 3 As shown, the brain choroid plexus volume prediction information may include at least one of the following: brain choroid plexus volume prediction value, brain choroid plexus volume high overload probability, and brain choroid plexus volume high overload risk level.

[0095] Example 4: Experimental Verification

[0096] To verify the feasibility of the method of this invention, sample data including brain magnetic resonance imaging (MRI) images and fundus images were used for model construction and evaluation. MRI images of the sample subjects were used to obtain the volume of the choroid plexus, and fundus images were used to extract quantitative features from the fundus images. Basic information included age, gender, and years of education.

[0097] In one embodiment, a total of 815 brain imaging samples were analyzed, of which 578 had fundus images. The volume of the choroid plexus was determined by segmenting the choroid plexus region in the T1-weighted images of brain magnetic resonance imaging. After retinal vessel segmentation and arteriovenous classification, the vascular morphology features of whole vessels, arteries, and veins were extracted from the fundus images. For ease of description, fractal dimension, vessel density, arterial fractal dimension, arterial vessel density, and average vein width are denoted as FD, VD, AFD, AVD, and VAW, respectively.

[0098] In one embodiment, a high-overload label for the choroid plexus volume is determined based on the distribution of choroid plexus volume. Specifically, samples whose choroid plexus volume is located in the quartile of the overall sample are designated as the high-overload target group, and the remaining samples are designated as the non-high-overload group. The choroid plexus volume prediction model takes quantitative features of fundus images and basic information as input and outputs the probability of high choroid plexus volume.

[0099] In one embodiment, FD, VD, AFD, AVD, and VAW are selected as quantitative features of fundus images, and the quantitative features of fundus images are combined with age, gender, and years of education as input data for a brain choroid plexus volume prediction model.

[0100] like Figure 4As shown, in one embodiment, quantitative features and basic information from fundus images are used as input, and a support vector machine model is employed to predict the upper quartile state of the choroid plexus volume. Model training and testing utilize hierarchical five-fold cross-validation. Results show that the choroid plexus volume prediction model has predictive ability for the upper quartile state of the choroid plexus volume, with an average receiver operating characteristic (ROC) area under the curve (AUC) of 0.80 ± 0.03.

[0101] The experimental results above demonstrate that the brain choroid plexus volume prediction model based on fundus images described in this invention can output prediction results corresponding to the brain choroid plexus volume obtained from brain magnetic resonance imaging (MRI) images, based on quantitative features and basic information from fundus images. This result can serve as auxiliary assessment information for the brain choroid plexus volume status, providing a reference for deciding whether to conduct further MRI examinations, follow-up assessments, or manual review.

[0102] Although specific embodiments, model types, experimental results, and drawings of the present invention have been disclosed for illustrative purposes to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that any substitution, adjustment, or combination of specific steps, threshold settings, model types, feature processing methods, image types, and module combinations without departing from the spirit and scope of the present invention and the claims should fall within the protection scope of the present invention.

Claims

1. A method for predicting the volume of the choroid plexus in the brain based on fundus images, characterized in that, Includes the following steps: 1) Obtain fundus images of the target subject; 2) The fundus images are processed to obtain model input data for predicting the volume of the choroid plexus in the brain; wherein the model input data includes at least one of the following: preprocessed fundus images, fundus image processing results, quantitative features of fundus images, and basic information; 3) Input the model input data into the pre-constructed brain choroid plexus volume prediction model to obtain the brain choroid plexus volume prediction information of the target object; The brain choroid plexus volume prediction model is constructed based on the mapping relationship between the fundus image data of the training subjects and the brain choroid plexus volume label, wherein the brain choroid plexus volume label is determined by the brain choroid plexus region in the brain magnetic resonance image of the training subjects. The brain choroid plexus volume prediction information includes at least one of the following: brain choroid plexus volume prediction value, standardized brain choroid plexus volume prediction value, brain choroid plexus volume high overload probability, and brain choroid plexus volume high overload risk level.

2. The method as described in claim 1, characterized in that, In step 1), the fundus image includes at least one of fundus photographic image, optical coherence tomography (OCT) angiography image, and OCT structural image.

3. The method as described in claim 2, characterized in that, Step 2) involves processing the fundus image, including at least one of the following steps: Image quality control is performed on the fundus images; The fundus image is preprocessed, and the preprocessing includes at least one of image normalization, brightness correction, contrast enhancement, noise suppression, and artifact correction; Based on the type of fundus image, at least one of the fundus structural regions, vascular regions, blood flow signal regions, and retinal layer structural regions in the fundus image is located, segmented, or extracted; When the fundus image is a fundus photograph image, the processing of the fundus image includes at least one of the following: optic disc region localization, macular region localization, retinal vascular region segmentation, vascular enhancement, vascular binarization, vascular skeletonization, and arteriovenous classification. When the fundus image is an optical coherence tomography angiography image, the processing of the fundus image includes at least one of blood flow signal extraction, vascular region segmentation, retinal or choroidal vascular layer segmentation, perfusion region extraction and avascular region extraction. When the fundus image is an optical coherence tomography (OCT) structural image, the processing of the fundus image includes at least one of retinal layer segmentation, choroidal region segmentation, macular region localization, and optic disc region localization.

4. The method as described in claim 1, characterized in that, In step 2), the model input data includes at least one of the following: preprocessed fundus image, fundus image processing result, quantitative features of fundus image, and basic information. The fundus image processing results include at least one of the following corresponding to the fundus image type: retinal vessel segmentation map, retinal artery segmentation map, retinal vein segmentation map, vascular skeleton map, blood flow signal map, vascular layer segmentation map, perfusion area map, avascular area map, retinal layer segmentation map, and choroidal region segmentation map.

5. The method as described in claim 4, characterized in that, The quantitative features of the fundus image include at least one of the following: vascular morphology features, blood perfusion features, layer structure features, and regional morphology features corresponding to the type of fundus image; The quantitative features of the fundus image corresponding to the fundus photograph include at least one of the following: fractal dimension, vascular density, vascular width, average vascular width, vascular tortuosity, vascular branching characteristics, and arteriovenous width ratio of at least one vascular region among the whole blood vessels, arteries, and veins. The quantitative features of the fundus image corresponding to the optical coherence tomography angiography image include at least one of the following: retinal vessel density, perfusion density, blood flow area, avascular area, blood flow defect area, blood flow defect ratio, superficial vascular plexus vessel density, deep vascular plexus vessel density, peridiscal capillary density, and choroidal capillary layer blood flow defect parameters. The quantitative features of the fundus image corresponding to the optical coherence tomography structural image include at least one of the following: retinal thickness, choroidal thickness, retinal nerve fiber layer thickness, ganglion cell complex thickness, macular thickness, macular volume, and optic disc-related structural parameters.

6. The method as described in claim 1, characterized in that, In step 2), the basic information includes at least one of age, gender, years of education, body mass index, blood pressure status, diabetes status, smoking status, and alcohol consumption status.

7. The method as described in claim 1, characterized in that, In step 3), the process of determining the brain choroid plexus volume label includes: acquiring the brain magnetic resonance image of the training subject; segmenting the brain choroid plexus region from the brain magnetic resonance image; calculating the brain choroid plexus volume based on the brain choroid plexus region to obtain the brain choroid plexus volume label.

8. The method as described in claim 7, characterized in that, The brain magnetic resonance imaging includes T1-weighted images; the brain choroid plexus volume label includes at least one of the original brain choroid plexus volume and the brain choroid plexus volume corrected for transcranial volume.

9. The method as described in claim 1, characterized in that, In step 3), the probability of high brain choroid plexus volume overload or the risk level of high brain choroid plexus volume overload is determined based on the high brain choroid plexus volume overload label, which is generated based on the relationship between the original brain choroid plexus volume of the training subject and a preset threshold. The preset threshold includes at least one of the quantile threshold of the original brain choroid plexus volume in the training set and the original brain choroid plexus volume threshold determined based on the normal reference population. The risk level of high choroid plexus volume overload includes at least two of the following: low risk, medium risk, and high risk.

10. The method as described in claim 1, characterized in that, In step 3), the construction process of the brain choroid plexus volume prediction model includes: Acquire fundus imaging data and brain magnetic resonance images of multiple training subjects; The original choroid plexus volume was determined based on the brain magnetic resonance images of the training subjects; The input data for the training model is determined based on the fundus image data of the training subjects; Using the training model input data as input, and at least one of the original choroid plexus volume, choroid plexus volume high-overload label, and choroid plexus volume high-overload risk level label as supervision information, the choroid plexus volume prediction model is constructed. The brain choroid plexus volume prediction model includes at least one of a machine learning model, a deep learning model, and a fusion model. The machine learning model includes at least one of the following: support vector machine, logistic regression, random forest, gradient boosting decision tree, LightGBM, XGBoost, CatBoost, multilayer perceptron, ridge regression, LASSO regression, and elastic network regression. The deep learning model includes at least one of the following: convolutional neural network, residual network, DenseNet, EfficientNet, visual Transformer, Swing Transformer, U-Net encoder network, attention neural network, and multimodal fusion neural network.

11. A brain choroid plexus volume prediction system based on fundus images, characterized in that, include: The image acquisition module is used to acquire fundus images of the target object; The image processing module is used to process the fundus image to obtain model input data for predicting the volume of the choroid plexus in the brain; wherein, the model input data includes at least one of the following: preprocessed fundus image, fundus image processing result, quantitative features of fundus image, and basic information; The prediction module is used to input the model input data into a pre-constructed brain choroid plexus volume prediction model to obtain brain choroid plexus volume prediction information of the target object; The result output module is used to output the predicted brain choroid plexus volume information; The brain choroid plexus volume prediction model is constructed based on the mapping relationship between the fundus image data of the training subjects and the brain choroid plexus volume labels. The brain choroid plexus volume label is determined by the brain choroid plexus region in the brain magnetic resonance image of the training subject; The brain choroid plexus volume prediction information includes at least one of the following: brain choroid plexus volume prediction value, brain choroid plexus volume high overload probability, and brain choroid plexus volume high overload risk level.

12. An electronic device and a computer-readable storage medium, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the brain choroid plexus volume prediction method based on fundus images as described in any one of claims 1 to 10. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the brain choroid plexus volume prediction method based on fundus images as described in any one of claims 1 to 10.

Images