Alzheimer disease risk assessment method and system based on eye-brain cross-modal AI

Abstract

The invention discloses an Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI, and relates to the technical field of brain disease prediction, and the method comprises the following steps: S1, data preprocessing: carrying out the standardization processing of an original OCTA image, and generating graph structure data; s2, a dual-path feature extraction mechanism: extracting local form anomalies and global topology anomalies of OCTA blood vessels in parallel through form paths and topology paths; s3, a cross-modal confrontation alignment mechanism is adopted, and cross-modal features are aligned and migrated by adopting the cross-modal confrontation alignment mechanism; and S4, an uncertainty quantification and hierarchical output mechanism: generating prediction distribution, calculating uncertainty based on the prediction distribution, and triggering hierarchical diagnosis and treatment suggestions. According to the Alzheimer disease risk assessment method and system based on the eye-brain cross-modal AI, accurate identification of AD pathology related capillary features can be effectively achieved, and clinical diagnosis is accurately assisted.

Description

Technical Field

[0001] This invention relates to the field of brain disease prediction technology, and in particular to a method and system for Alzheimer's disease risk assessment based on eye-brain cross-modal AI. Background Technology

[0002] Early diagnosis of Alzheimer's disease (AD) is crucial for slowing disease progression. Current technology utilizes optical coherence tomography (OCTA) for angiography. Researchers have discovered homology between retinal and brain microvessels, and their morphological changes are correlated with AD pathology, offering the possibility of non-invasive screening. Current technologies are mainly divided into two categories: 1. Single-modal deep learning model: Uses convolutional neural networks (CNNs) to perform end-to-end disease classification or regression on single-source (fundus only or brain only) OCTA images of humans or animals, and outputs a risk score.

[0003] 2. Statistical association analysis: Calculate the population statistical correlation between fundus vascular parameters (such as vascular density and fractal dimension) and brain AD pathological markers.

[0004] However, existing technologies have the following defects and shortcomings: (1) The "black box" nature of the model decision-making process: Most existing deep learning models adopt an end-to-end "holistic" architecture, directly mapping OCTA images to disease risk probabilities. Although this method performs well on specific datasets, it lacks interpretability. Doctors cannot know what features (such as sparse blood vessels, abnormal branching, or artifacts) the model is based on in making its judgments. In the medical field, the lack of transparency in decision-making directly leads to a lack of clinical trust. In addition, a single convolutional network is difficult to capture both local texture changes (morphology) and global network connection patterns (topology) at the same time, and it is easy to overlook subtle topological changes in early AD.

[0005] (2) The "Domain Shift" Problem in Cross-Modal Transfer Learning: Due to the extreme difficulty in obtaining early OCTA data of human AD with diagnostic labels, researchers often use transgenic animal models (such as 5xFAD or APP / PS1 mice) for pathological studies. However, there are significant differences in scale, distribution, and anatomical structure between mouse brain vessels, mouse retinal vessels, and human retinal vessels, which is known as "Domain Shift." Directly applying models trained on animal data to human clinical data often leads to a significant performance drop. Existing transfer learning methods mostly focus on fine-tuning and fail to fundamentally learn the "domain-invariant" essential disease characteristics across species and organs.

[0006] (3) Lack of a safety mechanism for quantifying uncertainty: Traditional deep neural networks are typically deterministic, meaning they output a definite prediction for any input (even severe artifacts or out-of-distribution data). This "blind confidence" is extremely dangerous in medical diagnosis. For example, when given a very poor-quality OCTA image, the model may incorrectly give a high-risk prediction due to a certain noise pattern, and the system itself cannot perceive or warn of this uncertainty. The lack of a self-evaluation mechanism limits the safe deployment of AI systems in real clinical environments. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for Alzheimer's disease risk assessment based on eye-brain cross-modal AI, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides an Alzheimer's disease risk assessment method based on eye-brain cross-modal AI, comprising the following steps: S1. Data preprocessing: Standardize the original OCTA image and generate graph structure data; S2. Dual-path feature extraction mechanism: Through morphological path and topological path, local morphological abnormalities and global topological abnormalities of OCTA vessels are extracted in parallel, overcoming the limitation that a single path cannot take into account the expression of two types of features, and realizing the comprehensive capture of pathological features. S3. Cross-modal adversarial alignment mechanism: To address cross-modal domain shifts between the source and target domains in fundus and brain regions, a cross-modal adversarial alignment mechanism is employed to align and transfer cross-modal features. S4. Uncertainty Quantification and Hierarchical Output Mechanism: By generating a predicted distribution and calculating uncertainty based on the predicted distribution, hierarchical diagnosis and treatment suggestions are triggered to solve the problem of blindly outputting high-confidence results under low-quality OCTA images in clinical practice.

[0009] Preferably, the specific steps of S1 are as follows: S11. Image Standardization: By using histogram equalization and intensity normalization, the pixel values ​​of the original OCTA image are mapped to the [0, 1] interval to obtain a standardized image. S12. Vascular skeleton extraction and attribute graph construction: Perform vascular segmentation on the standardized image and extract the vascular skeleton, converting the standardized image into graph structure data.

[0010] Preferably, the specific steps of S2 are as follows: S21. Morphological Path: A morphological feature extraction network is used to process standardized images and capture the morphological features of microvessels, including density reduction, texture abnormalities, and signal loss. S22. Topology Path: Utilize topology features to extract network processing graph structure data and capture topology features that are difficult to quantify by the human eye. Topology features include global network connection anomalies.

[0011] Preferably, the morphological feature extraction network in S21 employs asymmetric convolution.

[0012] Preferably, the topology feature extraction network in S22 is a graph neural network.

[0013] Preferably, the specific steps of S3 are as follows: S31, Cross-attention fusion and contribution generation: Deeply fuse the features of morphological path and topological path to generate unified features, and use the contribution generator sub-network to obtain the contribution weights of morphological path and topological path. S32. Cross-modal adversarial training: Achieve cross-modal transfer from animals to humans through adversarial domain adaptation strategies.

[0014] Preferably, the specific steps of S32 are as follows: S321. Construct and input the cross-modal training data stream; S322. Set up a modality discriminator to distinguish the modality source of the input features; S323. Set up a shared encoder to form an adversarial game with the modality discriminator, and generate features that make the modality discriminator unable to distinguish the source of the modality; S324. Construct a total loss function that retains common features related to AD pathology.

[0015] Preferably, the specific formula for the total loss function in S32 is as follows: ; in, L cls It is a classification loss based on the brain's gold standard label. y For real labels, The probability of AD risk predicted by the model. L adv It is the classification loss of the discriminator. d For real modal labels, The modal probabilities predicted by the discriminator. The weighting coefficients are used to balance the classification loss and the adversarial loss.

[0016] Preferably, the specific steps of S4 are as follows: S41. Predictive distribution generation: Using the Monte Carlo strategy, a set of predicted probabilities corresponding to OCTA blood vessels is generated through multiple independent sampling and inference. S42. Uncertainty Calculation: Based on the prediction distribution, calculate the predicted value and uncertainty score; S43. Based on the uncertainty calculation results, trigger the hierarchical diagnosis and treatment recommendations, which adopt a three-level decision-making logic.

[0017] A system for Alzheimer's disease risk assessment based on eye-brain cross-modal AI includes: Data acquisition and preprocessing module: used to acquire raw OCTA images, perform standardization processing, and generate morphological and topological graph structure output data; Dual-path feature extraction module: used to extract the morphological and topological features of blood vessels through a dual-path feature extraction mechanism; Cross-attention module: used to deeply fuse features from two paths to generate unified features; Cross-modal adversarial training module: used to achieve alignment and transfer of cross-modal features between the eye and brain using a cross-modal adversarial alignment mechanism; Uncertainty Quantification and Hierarchical Output Module: Used to quantify the degree of uncertainty in system predictions and output hierarchical diagnostic and treatment recommendations.

[0018] Therefore, the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI described above have the following beneficial effects: (1) This invention proposes an interpretable deep learning network structure specifically adapted to OCTA images. It adopts a dual-path architecture to extract vascular morphological features and topological features in parallel and independently, thereby achieving complementary enhancement of the two types of features and accurately capturing microvascular features related to AD pathology. It introduces a cross-modal adversarial alignment mechanism to successfully remove species-specific features, enabling the model to focus on the common expression of AD pathology and significantly improve generalization ability.

[0019] (2) The present invention achieved an AUC value of 0.865±0.04 in the AD pathological auxiliary diagnosis task, which showed better classification performance compared with the comparison model. It has high discrimination between AD and normal samples and good diagnostic efficacy. It can realize the accurate identification of AD pathological related microvascular features and can provide a reference for clinical auxiliary assessment.

[0020] (3) This invention significantly reduces the overall computational burden of the model, optimizes the number of parameters, adapts to the needs of rapid screening in outpatient clinics, has high inference efficiency, can be seamlessly deployed on existing hospital imaging workstations, does not require additional purchase of high-performance computing equipment, and has low memory usage; and has natural anti-interference ability against slight motion artifacts and uneven lighting noise during image acquisition, reducing the stringent requirements of clinical operation on image quality, improving the practicality and reliability of the system, and has strong robustness.

[0021] (4) This invention combines theoretical innovation with engineering feasibility. Through deep adaptation of technological innovation and clinical needs, it has significant economic benefits for clinical promotion and provides an efficient, accurate and feasible technical solution for AD pathological auxiliary diagnosis.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is an overall framework diagram of an embodiment of the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI of the present invention; Figure 2 This is a dual-path network representing an embodiment of the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI of the present invention; Figure 3 This invention relates to the feature fusion and parameter generation of an Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI. Figure 4 This is an adversarial training module for an embodiment of the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI of the present invention; Figure 5 This is a decision system for an embodiment of the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI of the present invention; Figure 6 These are mouse cerebral vascular OCT images from an embodiment of the Alzheimer's disease risk assessment method and system based on eye-brain cross-modal AI of the present invention. Figure 7 These are mouse fundus vascular OCT images from an embodiment of the Alzheimer's disease risk assessment method and system based on cross-modal AI of the eye and brain according to the present invention. Figure 8 This is an OCT image of the fundus vessels of a human eye, representing an embodiment of the Alzheimer's disease risk assessment method and system based on cross-modal AI of the eye and brain according to the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] Example Please see Figures 1-8 This invention provides a method and system for Alzheimer's disease risk assessment based on cross-modal AI of the eye and brain. The overall architecture design adopts a processing flow that combines serial and parallel processing, mainly including three stages: Data preprocessing and graph construction stage: The original OCTA image is standardized, and morphological image input and topological graph structure input are generated in parallel.

[0027] Dual-path feature extraction and adversarial training stage: morphological features and topological features are extracted using a dual-branch network, and adversarial training is performed through a cross-modal discriminator to achieve domain alignment of features.

[0028] Uncertainty Quantification and Clinical Decision-Making Stage: During reasoning, Monte Carlo Dropout is used to generate a predictive distribution, calculate the uncertainty score, and trigger tiered medical advice accordingly.

[0029] This invention provides an Alzheimer's disease risk assessment method based on eye-brain cross-modal AI. This method is a complete and systematic approach from model training to clinical inference, and its core workflow is as follows: Figure 1 As shown, the specific steps include: S1. Data preprocessing and graph structure generation.

[0030] S11, Image Standardization.

[0031] The input data for this step is the original OCTA scan image. To eliminate device differences (such as signal strength differences between Zeiss and Heidelberg devices), histogram equalization (CLAHE) and intensity normalization are first performed to map pixel values ​​to the [0, 1] interval, resulting in a standardized image I.

[0032] S12. Extraction of the vascular skeleton and construction of the attribute map.

[0033] To support topology-aware paths, the normalized image I is transformed into graph-structured data G(V, E). The specific steps are as follows: Segmentation and thinning: The U-Net network is used to segment blood vessels in the standardized image I, generating a binary blood vessel mask M. Subsequently, skeletonization is applied to extract the center lines of the blood vessels.

[0034] Node definition V: The bifurcation points and endpoints in the skeleton are defined as nodes in the graph. Each node v i Includes feature vector x v ∈ dv The feature vector includes features such as node coordinates (x, y) and local blood vessel diameter.

[0035] Edge definition E: connects any two nodes v i and v j The vascular segment is defined as edge e ij Each edge e ij Includes feature vector x e ∈ de The feature vector includes vessel segment length, tortuosity, average vessel diameter, etc.

[0036] By transforming visual images into highly abstract mathematical graph structures through S12, the foundation for graph neural network (GNN) processing is laid.

[0037] S2, Dual-path Interpretable Feature Extraction Network.

[0038] OCTA (Optical Characteristic Amplification) reveals both localized morphological abnormalities and global topological anomalies, which cannot be simultaneously expressed by a single pathway. Therefore, this invention employs a "dual-pathway" design, aiming to simulate the diagnostic thinking of an ophthalmologist who simultaneously observes "whether there is a reduction in blood vessels" (morphology) and "whether the vascular course is disordered" (topology). Figure 2 As shown, the specific steps are as follows: S21. Morphology-aware Path.

[0039] This pathway is responsible for capturing morphological features such as decreased microvascular density, abnormal texture, and signal loss.

[0040] Network backbone: adopts an improved ResNet-34 architecture.

[0041] Asymmetric Convolution Block: Considering that retinal vessels are mostly elongated structures, traditional 3×3 square convolution kernels are prone to introducing background noise and have low computational efficiency. This invention decomposes the standard convolution into cascaded or parallel structures of 3×1 and 1×3 asymmetric convolutions.

[0042] Mathematical expression: Let the input feature map be X, the convolution operation is defined as: ; in, Y To output the feature map, ReLU To modify the activation function of the linear unit, Conv 3×1 This is a vertical one-dimensional convolution operation. Conv 1×3 This is a horizontal one-dimensional convolution operation. BN It is a batch of normalized layers.

[0043] Asymmetric convolution decomposition technology not only enhances the network's sensitivity to vascular textures in both horizontal and vertical directions, but also significantly reduces the number of parameters and improves the anisotropic response capability of the receptive field. It balances the optimization of the receptive field and the number of parameters, making it more suitable for capturing the continuous features of microvessels.

[0044] Output: After global average pooling (GAP), the output is the morphological feature vector F. morph ∈ C1 .

[0045] S22. Topology-aware Path.

[0046] This approach utilizes graph neural networks (GNNs) to process the generated attribute graph G(V, E), aiming to capture global network connectivity anomalies that are difficult for the human eye to quantify.

[0047] Network selection: A Graph Attention Network (GAT) is used as the core component. GAT can dynamically assign weights based on the importance of neighboring nodes, which is crucial for analyzing key hubs in vascular networks (such as the bifurcation of main vessels).

[0048] Message passing mechanism: For node i, the feature update formula for its (l+1)th layer is: ; in, It is a non-linear activation function; W ( l ) is the first lThe learnable weight matrix of the layer is used to linearly transform the features of neighboring nodes; The attention coefficient represents the importance of node j to node i, and is usually calculated by the attention mechanism and normalized by Softmax. For nodes j In the The feature vector representation of a layer; N ( i ) represents a node i The set of neighboring nodes. This allows the model to focus on local subgraphs with topological anomalies (such as unusual branch angles or excessive curvature).

[0049] Output: Output the topological feature vector F through graph readout. topo ∈ C2 .

[0050] Graph neural networks process highly abstract graph-structured data (typically <500 nodes) rather than raw pixel-level images, reducing memory usage and allowing seamless deployment on existing hospital imaging workstations without the need for additional high-performance computing equipment. The synergistic design of asymmetric convolution and graph attention mechanisms provides natural resistance to minor motion artifacts and uneven lighting noise during image acquisition, reducing the stringent image quality requirements of clinical operations, improving system practicality and reliability, and exhibiting strong robustness.

[0051] S3, Feature Fusion and Cross-Modal Adversarial Training.

[0052] For the fundus and brain regions, due to cross-modal domain shifts between animals and humans, this invention employs a cross-modal adversarial training alignment mechanism to address cross-species domain shifts, such as... Figures 3-4 As shown, the specific steps are as follows: S31, Cross-attention fusion and contribution generation.

[0053] Features F of the two paths morph and F topo Instead of simple concatenation, it uses a cross-attention module for deep fusion to generate a unified feature F. fusion .

[0054] Contribution quantification: To achieve interpretability, the system introduces a contribution generator subnetwork. This subnetwork contains two parallel fully connected layers, which respectively quantize F... fusion Mapped to scalar score S morph and S topo Then, the contribution weights α and β are obtained by Softmax normalization: ; Where α is the contribution weight of the morphological path, β is the contribution weight of the topological path, and α+β=1, S morph S is the original score for the morphological path. topo represents the original score of the topological path, and e is a natural constant.

[0055] Clinical significance: α represents the contribution of morphological features to the final diagnosis, and β represents the contribution of topological features. This directly answers the doctor's question of "why make this judgment".

[0056] S32, Cross-modal adversarial training strategy.

[0057] To address the challenge of migrating from animal models (source domain) to human patients (target domain), this invention introduces Adversarial Domain Adaptation.

[0058] Data stream: Paired cerebral vascular OCTA (I) using AD transgenic mice brain ) and fundus OCTA (I eye () as training data.

[0059] Discriminator design: A modality discriminator D is introduced, the goal of which is to distinguish the input features F. fusion Is it from the brain or the eye?

[0060] Adversarial game: The goal of the shared encoder (dual-path network) is to generate features that the discriminator cannot distinguish from the source, that is, to "confuse" modal information and retain only common AD pathological features (such as microvascular degeneration caused by amyloid deposition).

[0061] The specific formula for the total loss function is as follows: ; in, L cls It is a classification loss based on the brain's gold standard label. y The labels are real (0 indicates health, 1 indicates AD). The probability of AD risk predicted by the model. L adv It is the classification loss of the discriminator. d The modal labels are real (0 represents the brain, 1 represents the eye). The modal probabilities predicted by the discriminator. To balance the weights of the classification loss and the adversarial loss, a gradient reversal layer (GRL) is used to implement this adversarial process.

[0062] Results: After training, the features extracted by the encoder will have "modal invariance", which will enable the vascular lesion patterns learned in the mouse brain-eye model to be effectively transferred to human fundus image analysis, overcoming the feature distribution shift caused by species differences.

[0063] S4. Uncertainty Quantification and Incremental Decision Making.

[0064] In clinical low-quality images, traditional models tend to blindly output high confidence. To prevent models from giving incorrect high-confidence predictions when encountering out-of-distribution (OOD) data or low-quality images, such as... Figure 5 As shown, this invention employs an uncertainty quantification and hierarchical output mechanism, as detailed below: Monte Carlo Dropout (MC Dropout): During the inference phase, the Dropout layer is kept on, and the same input image is forwarded T times (e.g., T=50) to obtain a set of predicted probabilities {P1, P2, ..., PT}.

[0065] Uncertainty calculation includes forecast calculation and uncertainty score calculation: Predicted value: ; in, P t The AD shared prediction probability is obtained from the t-th Monte Carlo sampling. Uncertainty score: .

[0066] in, std ( P t ) is the standard deviation function.

[0067] The hierarchical output mechanism employs a three-level decision-making logic, as follows: Security rejection: If U > high (High threshold) The system prompts "Insufficient image quality or blurred features, it is recommended to re-examine or refer to another doctor" to avoid misdiagnosis.

[0068] High-confidence diagnosis: If U < low (Low threshold) The system outputs a determined risk level and an explanatory report.

[0069] In-depth analysis: If low ≤ U ≤ highThe system then enters interactive analysis mode, prompting doctors to focus on the highlighted areas of the heat map and providing α and β reference values.

[0070] This invention, through the rational design of feature dimensions (512 dimensional features and 256 dimensional features) and a lightweight contribution generator (only two fully connected layers), enables a complete inference process including T=50 MCDropout samplings to be completed within 3 seconds on a standard clinical workstation (such as an NVIDIA T4 graphics card). This is suitable for the needs of rapid screening in outpatient clinics and has high inference efficiency.

[0071] S5. Experimental verification.

[0072] S51. Experiment setup and dataset construction.

[0073] Dataset Description: To verify cross-modal transfer capabilities, the following dataset was constructed in this embodiment: Source domain training set: contains paired brain and fundus OCTA images of 50 AD transgenic mice (e.g., 5xFAD model) and 50 wild-type (WT) mice, with labels determined by genotype.

[0074] Target domain test set: contains data from 100 human subjects, collected from multi-center clinical trials.

[0075] Grouping: 30 clinically diagnosed AD patients, 30 patients with mild cognitive impairment (MCI), and 40 cognitively normal control group (NC).

[0076] Gold standard: Confirmation via Amyloid-PET or cerebrospinal fluid biomarkers.

[0077] Image quality: Includes approximately 15% low-quality or interfering samples with coexisting ocular conditions (such as cataracts) to test uncertainty mechanisms.

[0078] Evaluation indicators: AUC (Area Under Curve): A core metric for measuring classification performance.

[0079] Sensitivity (Se) / Specificity (Sp): Measures the rate of missed diagnosis and the rate of misdiagnosis.

[0080] ARC (Accuracy-Rejection Curve): Measures the effectiveness of uncertainty quantification.

[0081] F1-Score: A comprehensive evaluation metric.

[0082] S52, Verification Results.

[0083] Result 1: Comparative analysis with existing technologies.

[0084] This step aims to demonstrate the superiority of the proposed dual-path network and cross-modal adversarial training over traditional single-modal CNNs and non-transfer models, as shown in Table 1 below.

[0085] Table 1. Performance comparison of different methods on human clinical datasets (AD vs. NC)

[0086] Results Analysis and Logical Derivation: Baseline Comparison: While the traditional ResNet-50 performs excellently on natural images, its AUC is only 0.812 on the OCTA microvessel classification task. This is mainly because CNNs struggle to explicitly capture the connectivity features of vascular networks and are prone to overfitting to noise (such as vascular shadows) in human eye data. The simple GNN model (AUC 0.795), although capturing the topological structure, loses microvessel density (VD) and texture information, demonstrating the limitations of single-modality models.

[0087] The necessity of transfer learning: directly applying a network trained on a mouse model to the human eye ("dual-path network (non-adversarial)") resulted in an AUC of only 0.624, even lower than random guessing. This strongly confirms the severity of "domain shift"—there are fundamental differences between mouse retinal vessels and human retinal vessels in terms of diameter and branching patterns.

[0088] In this embodiment, the AUC of the present invention on the human clinical test set can reach 0.865±0.04, which shows better classification performance than the comparison model. This result is the average result of multiple repeated experiments. The advantage is attributed to two factors: first, the dual-path architecture complements morphological and topological features; second, cross-modal adversarial training successfully removes species-specific features, allowing the model to focus on the common pathological features of AD (such as capillary thinning caused by amyloid deposition). This result strongly supports the technical effect of improving generalization ability.

[0089] Result 2: Dynamic evolution of interpretability weights (α and β).

[0090] A key feature of this invention is its ability to output contribution scalars α (morphology) and β (topology). To demonstrate its clinical significance, this embodiment statistically analyzes the average weight distribution of the model outputs at different disease stages, as shown in Table 2 below.

[0091] Table 2. Distribution of characteristic contributions at different disease stages

[0092] This result not only demonstrates the model's high accuracy but also its reasoning ability, which aligns with medical logic. It can automatically adjust its focus based on the progression of the disease, shifting from early-stage "structural disorder" to later-stage "tissue loss." This provides doctors with invaluable auxiliary diagnostic information, overcoming the shortcomings of traditional AI that "only know what but not why," and greatly enhancing the patent's practicality.

[0093] Result 3: Uncertainty Quantification and Security Decision-Making Mechanism.

[0094] To verify the effectiveness of the "safe rejection" mechanism, this embodiment plots an accuracy-rejection curve (ARC). By progressively removing samples with the highest uncertainty score $U$, the changes in diagnostic accuracy of the remaining samples are observed.

[0095] Table 3. System Performance Improvement Based on Uncertainty Filtering (Accuracy-Rejection Analysis)

[0096] Analysis: Data shows that as the uncertainty threshold tightens (rejection rate increases), the system's accuracy exhibits a monotonically increasing trend. In particular, after rejecting the top 20% of high-uncertainty samples, the accuracy jumped from 82.8% to 89.7%. This indicates that the system can accurately identify cases it "cannot handle" (such as severe cataract obstruction or very early, weak features) and proactively hand them over to doctors, thus avoiding potential medical accidents. This mechanism is a key safety feature that distinguishes this invention from most competing products (which only pursue AUC).

[0097] This invention also provides a system for an Alzheimer's disease risk assessment method based on eye-brain cross-modal AI, comprising: Data acquisition and preprocessing module: used to acquire raw OCTA images, perform standardization processing, and generate morphological and topological graph structure output data; Dual-path feature extraction module: used to extract the morphological and topological features of blood vessels through a dual-path feature extraction mechanism; Cross-attention module: used to deeply fuse features from two paths to generate unified features; Cross-modal adversarial training module: used to achieve alignment and transfer of cross-modal features between the eye and brain using a cross-modal adversarial alignment mechanism; Uncertainty Quantification and Hierarchical Output Module: Used to quantify the degree of uncertainty in system predictions and output hierarchical diagnostic and treatment recommendations.

[0098] Example 1 The specific construction of the dual-path network: the morphological perception path (CNN branch).

[0099] A variant based on ResNet-34 is used. To accommodate the fine structure of retinal microvessels, this embodiment replaces the first 7×7 convolution of the original ResNet with a "Vessel Stem" module containing three 3×3 convolutions to reduce information loss due to downsampling.

[0100] The key improvement lies within the residual block. For the input feature map X, this embodiment introduces parallel asymmetric convolution branches: ; Factorized convolution not only reduces the number of parameters by 33%, significantly reducing the overall computational burden of the model and optimizing the number of parameters, but also forces the network to learn horizontal and vertical features, which is highly consistent with the growth characteristics of the retinal vascular network.

[0101] Topology-aware path (GNN branch): The input is a pre-constructed attribute graph. A three-layer graph attention layer (GAT layer) is used.

[0102] First layer: Input dimension D in =16 (node ​​features), output dimension D hidden =64, attention head count K=4.

[0103] Second layer: Output dimension D hidden =64, attention head count K=4.

[0104] Third layer: Output dimension D out =128, attention head number K=1. In order to extract layer-level features, this embodiment does not use simple average pooling, but uses hierarchical pooling (SAGPool), which removes 50% of nodes with low response values ​​after each layer, simulating the process of the human eye "refining" and focusing on key lesion areas.

[0105] Example 2 The specific operation of cross-modal adversarial training.

[0106] The training process is divided into two phases: Pre-training phase: Self-supervised pre-training of CNN branches (such as SimCLR) is performed using a large number of unlabeled human fundus images to learn general fundus feature representations.

[0107] Adaptation phase: Input: Source domain data (Paired mouse data), target domain data (Unlabeled human eye data).

[0108] Network structure: Shared feature extractor E, disease classifier C, domain discriminator D.

[0109] Forward propagation: E extracts feature f. C predicts AD risk based on f. D predicts the source of the feature based on f (whether it is Mouse-Brain, Mouse-Eye, or Human-Eye).

[0110] Optimization goal: Minimize the classification error of C in the source domain.

[0111] Minimize the discrimination error of D (train D).

[0112] Maximize the discrimination error of D (training E). This is achieved through a gradient reversal layer (GRL), which multiplies the gradient of D propagated back by −λ during backpropagation.

[0113] Through this game, the final feature f can distinguish between AD and cannot distinguish the species origin, thus realizing cross-modal knowledge transfer.

[0114] Example 3 Clinical workflow integration.

[0115] The system is deployed on ophthalmology workstations in hospitals: Data collection: The technician used an OCTA device to scan the patient's macular area (3x3mm).

[0116] Inference: The system automatically reads DICOM data and performs standardized preprocessing.

[0117] Calculation: Start MC Dropout (T=50) to calculate the mean risk probability μ and uncertainty U in parallel.

[0118] Report generation: The prediction results are displayed, for example, "AD Risk Index: 0.85 (High Risk)".

[0119] Explanatory indicators are displayed, such as "Judgment criteria: morphological abnormalities weighted at 70% (mainly due to decreased density of deep vessels), and topological abnormalities weighted at 30%".

[0120] Display heatmap: Overlay the lesion area (e.g., use Grad-CAM technology to highlight the ischemic area).

[0121] Safety Tip: If U > 0.15, mark "low confidence" in a prominent color and it is recommended to review the results using the MMSE scale.

[0122] Therefore, the Alzheimer's disease risk assessment method and system based on cross-modal AI of the eye and brain adopted in this invention has high discrimination and good diagnostic efficacy between AD and normal samples. It can effectively realize the accurate identification of microvascular features related to AD pathology and can provide a reference for clinical auxiliary assessment.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An Alzheimer's disease risk assessment method based on eye-brain cross-modal AI, characterized in that, Includes the following steps: S1. Data preprocessing: Standardize the original OCTA image and generate graph structure data; S2. Dual-path feature extraction mechanism: Through morphological path and topological path, local morphological abnormalities and global topological abnormalities of OCTA vessels are extracted in parallel, overcoming the limitation that a single path cannot take into account the expression of two types of features, and realizing the comprehensive capture of pathological features. S3. Cross-modal adversarial alignment mechanism: To address cross-modal domain shifts between the source and target domains in fundus and brain regions, a cross-modal adversarial alignment mechanism is employed to align and transfer cross-modal features. S4. Uncertainty Quantification and Hierarchical Output Mechanism: By generating a predicted distribution and calculating uncertainty based on the predicted distribution, hierarchical diagnosis and treatment suggestions are triggered to solve the problem of blindly outputting high-confidence results under low-quality OCTA images in clinical practice.

2. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Image Standardization: By using histogram equalization and intensity normalization, the pixel values ​​of the original OCTA image are mapped to the [0, 1] interval to obtain a standardized image. S12. Vascular skeleton extraction and attribute graph construction: Perform vascular segmentation on the standardized image and extract the vascular skeleton, converting the standardized image into graph structure data.

3. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 2, characterized in that, The specific steps of S2 are as follows: S21. Morphological Path: A morphological feature extraction network is used to process standardized images and capture the morphological features of microvessels, including density reduction, texture abnormalities, and signal loss. S22. Topology Path: Utilize topology features to extract network processing graph structure data and capture topology features that are difficult to quantify by the human eye. Topology features include global network connection anomalies.

4. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 3, characterized in that, The morphological feature extraction network in S21 employs asymmetric convolution.

5. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 4, characterized in that, The topology feature extraction network in S22 is a graph neural network.

6. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 5, characterized in that, The specific steps of S3 are as follows: S31, Cross-attention fusion and contribution generation: Deeply fuse the features of morphological path and topological path to generate unified features, and use the contribution generator sub-network to obtain the contribution weights of morphological path and topological path. S32. Cross-modal adversarial training: Achieve cross-modal transfer from animals to humans through adversarial domain adaptation strategies.

7. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 6, characterized in that, The specific steps of S32 are as follows: S321. Construct and input the cross-modal training data stream; S322. Set up a modality discriminator to distinguish the modality source of the input features; S323. Set up a shared encoder to form an adversarial game with the modality discriminator, and generate features that make the modality discriminator unable to distinguish the source of the modality; S324. Construct a total loss function that retains common features related to AD pathology.

8. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 7, characterized in that, The specific formula for the total loss function in S32 is as follows: ； in, L cls It is a classification loss based on the brain's gold standard label. y For real labels, The probability of AD risk predicted by the model. L adv It is the classification loss of the discriminator. d For real modal labels, The modal probabilities predicted by the discriminator. The weighting coefficients are used to balance the classification loss and the adversarial loss.

9. The Alzheimer's disease risk assessment method based on eye-brain cross-modal AI according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Predictive distribution generation: Using the Monte Carlo strategy, a set of predicted probabilities corresponding to OCTA blood vessels is generated through multiple independent sampling and inference. S42. Uncertainty Calculation: Based on the prediction distribution, calculate the predicted value and uncertainty score; S43. Based on the uncertainty calculation results, trigger the hierarchical diagnosis and treatment recommendations, which adopt a three-level decision-making logic.

10. The system for Alzheimer's disease risk assessment based on eye-brain cross-modal AI according to any one of claims 1-9, characterized in that, include: Data acquisition and preprocessing module: used to acquire raw OCTA images, perform standardization processing, and generate morphological and topological graph structure output data; Dual-path feature extraction module: used to extract the morphological and topological features of blood vessels through a dual-path feature extraction mechanism; Cross-attention module: used to deeply fuse features from two paths to generate unified features; Cross-modal adversarial training module: used to achieve alignment and transfer of cross-modal features between the eye and brain using a cross-modal adversarial alignment mechanism; Uncertainty Quantification and Hierarchical Output Module: Used to quantify the degree of uncertainty in system predictions and output hierarchical diagnostic and treatment recommendations.

Images