A multi-indicator analysis system for predicting the risk of retinal aneurysm rupture

Abstract

This invention relates to the field of medical image analysis, providing a multi-indicator analysis system for predicting the risk of retinal aneurysm rupture. The system includes: collecting patient-related structural image information and non-image-related medical indicator data; analyzing the joint variation relationship between the non-image-related medical indicator data and the corresponding structural image information over time based on historical sample data; combining the response mapping mechanism with the fluctuation of the current patient's non-image-related medical indicator data; performing feature encoding on the structural image processed by the image modulation module to extract image feature vectors reflecting the local tissue state; receiving the image feature vectors and performing pattern recognition to determine the rupture risk status of the retinal aneurysm. This system can improve the accuracy of early prediction of retinal aneurysm rupture risk.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis, and specifically relates to a multi-index analysis system for predicting the risk of rupture of retinal aneurysms. Background Technology

[0002] In the clinical prediction of the risk of retinal aneurysm rupture, structural images have long served as the primary source of information, forming the basis of diagnosis. However, with further research, it has become increasingly clear that relying solely on structural images for early risk identification faces numerous challenges. First, in the early stages of retinal aneurysm formation, structural abnormalities often manifest as subtle irregularities in the vessel wall, changes in local permeability, or mild morphological distortions. These changes are extremely subtle and difficult to accurately distinguish at conventional image resolution. Furthermore, the complex textures of background tissues and visual noise further mask these early features, significantly reducing the diagnostic efficacy of the images.

[0003] To compensate for the limitations of image representation, some existing research attempts to introduce non-image-based medical data, such as blood pressure, blood glucose, and inflammatory factors, to provide supplementary diagnostic criteria. Common methods include feature concatenation of image features and non-image indicators, and the introduction of multimodal data during model input. However, most of these methods are shallow fusions, lacking modeling of the deep logical relationships between image and non-image data, making it difficult to achieve targeted guidance of indicators on image representation. Furthermore, many existing methods fail to uncover the statistical correlation between non-image indicators and image region changes from historical samples, and have not established mechanisms for non-image data to drive image regulation, resulting in limited fusion effectiveness.

[0004] More importantly, currently widely used image convolution-based feature extraction methods typically rely on the image's inherent sharpness and edge discernibility. If the target region in the input image is not adequately represented, even a complex model structure will struggle to capture genuine risk signs. In clinical practice, systemic physiological changes in patients often precede visible structural damage in images, a process that current image processing methods have not fully utilized.

[0005] Therefore, a fundamental flaw in existing technologies lies in the lack of a mechanism to regulate image expression by using non-image indicators as the dominant factor. This makes it impossible to capture potential risk signals in advance before significant morphological changes occur in the image, thus limiting the predictive value of images in the early stages of aneurysm rupture risk. Summary of the Invention

[0006] To address the problems in the prior art, this invention provides a multi-index analysis system for predicting the risk of retinal aneurysm rupture, comprising:

[0007] The data acquisition module is used to collect patient-related structural image information and non-image medical indicator data. The structural image information is used to present the state of retinal vessels and surrounding tissues. The non-image medical indicator data includes time-series parameters that can characterize the patient's circulatory metabolic state and / or vascular stress changes and / or inflammatory response trends.

[0008] The fluctuation correlation module is used to analyze the joint change relationship between the non-image-type medical indicator data and the corresponding structural image information in the time dimension based on historical sample data, and to establish a response mapping mechanism to reflect the fluctuation correlation pattern between the two types of data.

[0009] The image modulation module is used to perform expression enhancement or expression suppression operations on regions in the structural image that are sensitive to fluctuations, based on the fluctuations of the current patient's non-image-related medical indicator data and in conjunction with the response mapping mechanism. The modulation operations are used to adjust the image response intensity distribution of the target region to reflect its potential risk importance under the current indicator state.

[0010] The image feature encoding module is used to encode the features of the structural image after it has been processed by the image control module, and to extract image feature vectors that reflect the local tissue state.

[0011] The risk identification module is used to receive the image feature vector and perform pattern recognition to determine the rupture risk status of the retinal aortic aneurysm.

[0012] Furthermore, the non-image-based medical indicator data includes one or more of the following: blood pressure, heart rate variability, blood glucose level, blood lipid composition, and C-reactive protein level.

[0013] Furthermore, the structural image information includes the ganglion cell layer, the nuclear layer, and the outer reticular layer.

[0014] Furthermore, the analysis of the joint changes in the non-image-related medical indicator data and the corresponding structural image information over time, based on historical sample data, includes:

[0015] Time standardization processing was performed on non-image-based medical indicator data;

[0016] Extract local fluctuation features of non-image-related metrics within each sliding time window;

[0017] Extract structural image information corresponding to the sliding time window;

[0018] The structural image is divided into several regions based on anatomical structure or image segmentation algorithms;

[0019] Extract preset image features from each divided region;

[0020] The temporal correlation between different fluctuation feature dimensions and the features of each image region was calculated based on the Pearson correlation coefficient, and a statistical response map was established to reflect the degree of response of the fluctuation pattern to changes in image structure.

[0021] Construct a heterogeneous graph structure, where the set of nodes includes fluctuating feature nodes and image response nodes, and edges represent potential response paths between the two.

[0022] By using a graph neural network propagation mechanism, the connection weights between nodes are trained end-to-end, and an attention mechanism is introduced to assign the influence intensity of different indicators on image regions.

[0023] Furthermore, the analysis of the joint changes in the non-image-related medical indicator data and the corresponding structural image information over time, based on historical sample data, includes:

[0024] Construct a potential propagation response field based on the image region map, so that the fluctuation of each index diffuses on the image structure map in the form of Gaussian kernel weights, and dynamically forms a weight gradient distribution;

[0025] For a non-image indicator in historical samples, the fluctuation characteristics are extracted within a sliding time window, and its dynamic fluctuation potential is extracted through a gating unit.

[0026] The structural image is divided into a region map, where nodes represent local regions in the image and edges represent the spatial connections between regions.

[0027] Construct a potential response weight field on the graph, where the response weight of each node is defined by a Gaussian diffusion function;

[0028] The response fields of multiple indicators are synthesized into an overall response field according to a weighted average or maximization mechanism.

[0029] Furthermore, the image control module specifically includes:

[0030] Perform two-dimensional pixel grayscale conversion on structural image information;

[0031] A weight matrix is ​​generated based on a fluctuation response mapping mechanism.

[0032] The image of the corresponding region is processed and adjusted according to the weight matrix.

[0033] Furthermore, a response transition band is introduced at the region boundary, and a spatial bilateral filtering algorithm is used to smooth the adjustment factor, so that the image response intensity changes continuously in space.

[0034] Furthermore, control operations will only be performed when the regional response intensity is greater than or less than the preset value.

[0035] Furthermore, the image feature encoding module adopts a dual-branch structure and introduces residual connection paths. One branch encodes the overall structural layout and background configuration of the image, while the other branch focuses on the high-response regions identified by the image modulation module, generating local and global feature maps respectively through convolutional pooling paths.

[0036] Furthermore, the risk identification module adopts a parallel dual-channel structure, with one channel used to extract the global pattern of the overall image structure and the other channel focusing on the feature information of the high-response region after modulation.

[0037] The system dynamically adjusts the importance of the output features of the two channels through an attention mechanism and performs feature fusion in the final output stage to improve the sensitivity to discriminative changes in key regions.

[0038] This invention, by constructing a fluctuation response mapping mechanism, achieves dynamic regulation of the expressive content of structural images using non-image-based medical indicator data, overcoming the deficiency of insufficient expressive power of image features in the early stages of traditional methods. The system can identify risk-sensitive regions in images based on individual physiological fluctuations and implement expression enhancement or suppression operations, enabling structural images to reflect potential risk signs even before significant pathological morphological changes occur.

[0039] By introducing a multimodal data fusion and regulation mechanism, this invention effectively improves the pathological relevance of image features and enhances the discriminative dimensional representation of rupture risk in image feature vectors. This design approach, which guides image structure reconstruction with non-image data, significantly enhances the robustness and adaptability of the model in handling complex individual differences, subtle early symptoms, and blurred images, among other clinical challenges.

[0040] Furthermore, the system architecture of this invention has good modular scalability, can adapt to different types of image acquisition devices and indicator data input sources, supports personalized parameter configuration and feature modeling strategies, and has high engineering practicality and clinical deployment potential. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0043] To more clearly illustrate the technical content of the present invention, the following detailed description of the multi-index analysis system for predicting the risk of retinal aneurysm rupture proposed in this invention is provided in conjunction with specific embodiments.

[0044] In clinical practice, early signs of aneurysms in retinal images often manifest as subtle structural changes or vascular abnormalities. These features are insufficiently expressed in conventional images, have unclear boundaries, and are easily affected by background interference, resulting in low accuracy of image-based early predictions. On the other hand, a patient's physiological state is usually accompanied by fluctuations in non-image-related indicators during changes in aneurysm risk, such as blood pressure, metabolism, and inflammation parameters. This information is not directly reflected in structural images. Existing methods fail to effectively utilize these non-image indicators to regulate image representation and lack mechanisms to mine the correlation between image and non-image data from historical samples. This results in insufficient risk perception of image data before it is input into the model, limiting the practicality and generalization ability of early predictions.

[0045] To address this, the present invention proposes a multi-indicator analysis system for predicting the risk of retinal aneurysm rupture. This system constructs an image preprocessing path oriented towards individualized indicator states. Before image feature extraction, based on the changing trends of non-image indicators in the current patient, it identifies structural regions in the image that are sensitive to these changes, and performs expression enhancement or suppression within the corresponding regions. This improves the image's ability to express risk at potential lesions, enabling subsequent identification models to more accurately capture early signs.

[0046] like Figure 1 As shown, the system includes a data acquisition module, a fluctuation correlation module, an image control module, an image feature encoding module, and a risk identification module. The system operates in the sequence of medical data acquisition—modeling—control—encoding—judgment. Data flows between modules are transmitted through structured interfaces, forming a closed-loop data processing path. Through this path, the system can complete the entire process from multi-source medical data of patients to rupture risk prediction without external intervention, significantly improving the ability to express image information and providing a foundation for early risk identification of retinal aortic aneurysms.

[0047] The entire system employs a clearly defined input, processing, and output workflow. It does not rely on manually set rules but automatically constructs an image enhancement mechanism that adapts to changes in individual characteristics by leveraging statistical learning relationships between images and non-image indicators in historical samples. This provides the predictive model with a more risk-aware input data foundation. The specific structure and data processing methods of each module in the system will be detailed in subsequent sections.

[0048] In the early stages of retinal aneurysm rupture risk, abnormal features on structural images often manifest as subtle morphological changes or localized texture interference, making them difficult to identify directly using conventional image processing methods. Simultaneously, the patient's overall physiological state is usually closely related to the formation and development of the aneurysm, especially under conditions of hemodynamic fluctuations, metabolic disorders, or inflammatory activation, which can cause microscopic stress changes in retinal vessels, the results of which may be indirectly reflected in the images. Therefore, relying solely on structural images for risk prediction is insufficient to account for individualized background differences. It is necessary to introduce non-image-related medical indicators closely related to image changes and to achieve more effective risk identification through joint modeling based on these data. To this end, a data acquisition module is set up to collect patient-related structural image information and non-image-related medical indicator data. The structural image information is used to present the state of retinal vessels and surrounding tissues, while the non-image-related medical indicator data includes temporal parameters that characterize the patient's circulatory metabolic state and / or vascular stress changes and / or inflammatory response trends.

[0049] Structural image information refers to digital image information acquired by medical imaging equipment that characterizes the tissue structure, blood vessel distribution, and inter-layer configuration within the patient's retinal region. This information is stored in matrix form and includes parameters such as grayscale values, inter-layer contrast, and vascular channels, and typically originates from optical coherence tomography (OCT) or vascular imaging equipment.

[0050] Non-image-based medical indicator data refers to time-series numerical data that is not directly presented in image form but can be obtained through medical monitoring, laboratory testing, electronic medical records, and other means. This data primarily reflects a patient's current and historical characteristics of blood circulation, metabolic regulation, endocrine and inflammatory states, including but not limited to blood pressure, heart rate variability, blood glucose levels, blood lipid components, and C-reactive protein levels.

[0051] The data acquisition module includes an image acquisition submodule and an indicator acquisition submodule. The image acquisition submodule interfaces with a standard medical imaging interface to access structural image data from optical imaging equipment at specific locations and under specific modes. After performing image layering and channel preprocessing, it forms a standard image data structure. The indicator acquisition submodule interfaces with the hospital information system, monitoring equipment, or historical electronic health records. It automatically extracts time-series indicator data matching the target patient and performs missing value imputation, unit standardization, and time alignment to make it suitable for subsequent joint modeling tasks.

[0052] In a preferred implementation, the image acquisition submodule uses an OCTA device with high resolution and vascular enhancement capabilities to scan and acquire complete retinal layer-level image data from the fovea to the optic disc region. The images may include information from multiple layers, such as the ganglion cell layer, the inner nuclear layer, and the outer reticular layer, to fully capture structural states such as vascular perforation, dilation, and rupture. The indicator acquisition submodule is equipped with a physiological state synchronization control unit to ensure the consistency of the time window between image acquisition and indicator acquisition, and to filter high-frequency interference components in the indicator data to improve the stability of subsequent related modeling.

[0053] In another preferred implementation, the data acquisition module includes a data consistency detection mechanism to check the consistency between indicator data and image data in terms of patient identity, time labels, and acquisition conditions. If inconsistencies are detected, a data retrieval mechanism is triggered to prevent data structure errors from interfering with subsequent models.

[0054] Through the above settings, the data acquisition module can achieve collaborative collection and structured organization of data from different modalities of the same patient, ensuring the alignment of the collected data in the time dimension and the complementarity in the content dimension, and providing a unified data foundation for cross-modal analysis of downstream modules.

[0055] This module enables the system to fully access both image and non-image information in the early stages and complete standardized input through a clearly structured data interface. This avoids the information silo problem caused by the single-source image input in traditional systems and helps improve the sensitivity and stability of subsequent models in identifying potential risks.

[0056] In a clinical study, a high-risk group was selected as the sample source. First, retinal structural images were acquired using OCTA (Optical Character Amplification), with a resolution of 512×512 pixels, covering the macula and its surrounding area. Subsequently, blood pressure, blood glucose, LDL-C, CRP, and other indicators from the past three months were simultaneously extracted from the hospital database to construct a complete patient data profile.

[0057] Early prediction of the risk of retinal aneurysm rupture largely depends on the accurate identification of subtle changes in vascular regions in images. However, such structural changes are often difficult to detect directly in the early stages due to their morphological characteristics. In contrast, non-image-related medical indicators such as blood pressure, blood lipids, and inflammatory response parameters often exhibit significant fluctuations before structural changes occur, characterizing the patient's current systemic stress response and pathological trends. Therefore, if the joint variation relationship between non-image-related medical indicators and structural image information in the temporal dimension can be effectively identified, a response mapping mechanism with indicators leading the image can be constructed to improve the controllability and risk specificity of image region changes. To this end, a fluctuation correlation module is set up to analyze the joint variation relationship between the non-image-related medical indicator data and the corresponding structural image information in the temporal dimension based on historical sample data, and to establish a response mapping mechanism to reflect the fluctuation correlation pattern between the two types of data.

[0058] The historical sample data refers to a multi-time period data set containing patient structural image information and non-image medical indicators. Each record unit contains at least a structural image and multiple time-series indicator parameters, and has a clear timestamp identifier.

[0059] This module implements the fluctuation response mapping mechanism in the following ways. First, the system performs time standardization on the non-image-related medical indicator data in the historical samples. Since non-image-related indicators are usually derived from clinical examinations, monitoring, or laboratory tests, their collection frequency and time distribution may be non-uniform, and some samples may also have missing data within a certain time period. To eliminate the time dimension differences between different samples, the system constructs a reference time axis based on a unified time baseline and performs time alignment on all indicator data. The time alignment process includes two steps: first, linear interpolation is performed on all indicator sequences to fill in missing time points; second, indicator data with inconsistent original sampling frequencies are transformed into data sequences with equal time intervals through resampling, so that each indicator dimension has a unified timestamp identifier under the same time axis, which facilitates subsequent joint modeling.

[0060] After time standardization, the system extracts local fluctuation features of non-image-related metrics within each sliding time window. Specifically, let the window size be... The original indicator sequence within the time window is represented as ,in Indicates the first The index value of the sequence, Representing the initial sequence, the system calculates the following statistical characteristics on this sequence to form a fluctuation feature vector. :

[0061] First-order difference feature, denoted as This reflects the trend of indicator changes;

[0062] Mean characteristic, expressed as This indicates the average state of the indicator within that window;

[0063] The standard deviation characteristic is expressed as ,in This represents the mean within the window, used to reflect the fluctuation range of the indicator;

[0064] Skewness and kurtosis are used to characterize the degree of deviation and steepness of the index distribution;

[0065] Extreme value difference or maximum-minimum ratio is used to reflect abnormal fluctuations or sudden changes.

[0066] This fluctuation feature vector serves as a dynamic behavior description of non-image-based medical indicators within the current time period.

[0067] Simultaneously, the system extracts structural image information corresponding to the time window. Since retinal structural images typically exhibit spatial partitioning, the system divides the structural image into several regions based on anatomical structures or image segmentation algorithms. These regions can be regular grid blocks, vascular distribution areas, arterial regions, or nerve fiber layers, etc. Within each partitioned region, the system extracts the following image features to construct the image response vector:

[0068] The mean and variance of grayscale reflect local brightness and contrast;

[0069] Edge intensity and edge density indicate the clarity of blood vessel boundaries;

[0070] Vascular permeability characteristics, including changes in vascular thickness, local light transmittance, and reflectance signal;

[0071] Local texture directionality or Gabor filter response is used to represent the trend of structural texture changes.

[0072] To achieve effective correlation between non-image metrics and image features, the system constructs a cross-modal temporal joint modeling mechanism. In one optional implementation, the system calculates the temporal correlation between different fluctuation feature dimensions and features of each image region based on the Pearson correlation coefficient, and establishes a statistical response map to reflect the degree of response of fluctuation patterns to changes in image structure.

[0073] In a preferred implementation, the system introduces a graph-structured attention network to model the response relationship between non-image metrics and image regions at a finer granularity. The system first constructs a heterogeneous graph structure, where the node set includes fluctuating feature nodes and image response nodes, with edges representing potential response paths between them. Through a graph neural network propagation mechanism, the connection weights between nodes are trained end-to-end. The system then introduces an attention mechanism to assign the influence intensity of different metrics on image regions. For a given fluctuating feature node… With image response node Its correlation weight It can be given by the following expression:

[0074]

[0075] in, and These are the transformation matrices for wave characteristics and image characteristics, respectively. This represents vector concatenation. This represents the attention parameter vector. Through the training process, the system can automatically discover non-image indicators with potential regulatory effects and quantify their influence on image regions in the form of weights.

[0076] The output of the aforementioned weight matrix is ​​the result of the fluctuation response mapping mechanism, used to represent the response regulation relationship of each non-image feature dimension to each region in the structural image during the current time period. Subsequently, based on this weight, the system enhances or suppresses the image region in the image regulation module to improve the recognizability of key structures in the image and assist in risk prediction.

[0077] In a preferred embodiment, the fluctuation correlation module introduces a structured latent fluctuation response field modeling mechanism. This mechanism differs from traditional graph neural networks or correlation modeling methods by innovatively constructing an explicit, dynamically adjustable fluctuation-response field to characterize the potential response trends of non-image-related medical indicators to local regions of structured images under temporal variations. The core idea of ​​this mechanism is to understand the changes in non-image modal indicators as a "perturbation source" in the image space. Through propagation modeling of the latent response field, the fluctuations of each indicator possess a spatial response trajectory, which is then mapped, updated, and output in a structured manner, thereby improving the accuracy of response matching and the interpretability of regulation among multiple modalities.

[0078] In medical data, non-image indicators may exhibit time-leading or multi-level responsiveness, and there are spatial coupling relationships between image structural regions. Single-point mapping will ignore its spatial propagation characteristics. At the same time, the effects of fluctuations may not be immediately apparent in the image, but will be gradually propagated through some structural path.

[0079] The structured latent fluctuation response field mechanism proposed in this implementation treats non-image indicators as dynamic source points, constructing a latent propagation response field based on the image region map. This allows fluctuations of each indicator to diffuse across the image structure map in the form of Gaussian kernel weights, dynamically forming a weight gradient distribution. This weight field is not a simple static mapping, but rather the result of the combined action of three constraint mechanisms: structure-driven, time-driven, and semantic-driven, possessing structural continuity, response adjustability, and temporal leading characteristics.

[0080] The system first performs time-series modeling on non-image indicators. Let's assume a certain non-image indicator in the patient's historical samples is within a sliding time window. The value sequence inside is The system extracts features such as its first derivative, local variation rate, trend direction, and acceleration, and extracts its dynamic fluctuation potential through a gating unit. ,Right now:

[0081]

[0082] in, , , , For weight parameters, For bias terms, This is the Sigmoid activation function.

[0083] Subsequently, the system divides the structural image into region maps. , where the set of nodes Represents a local region in an image, a set of edges. This represents the spatial connectivity between regions. The system constructs a potential response weight field on this map. The response weight of each node is defined by the following Gaussian diffusion function:

[0084]

[0085] in, The source node is selected based on the type of indicator (such as the optic disc center, macula, arterial area, etc.). This represents the structural distance between the source node and the target region. In response to the propagation scale factor, training optimization can be performed based on the pathological type.

[0086] The potential response field This corresponds to the distribution influence spectrum of a single indicator on the image structure map. The response fields of multiple indicators can be synthesized into an overall response field using a weighted average or maximization mechanism. ,in The importance weights for the regulation of indicators can be obtained through data-driven training.

[0087] Finally, the system will respond to the field. The input image modulation module serves as the basis for adjusting the expression intensity of each image region, performing enhancement or suppression operations. Specifically, for each image region... Its final adjustment coefficient is:

[0088]

[0089] in, For regional response intensity, In response to the field average, This is the amplification factor, used to control the amplification amplitude.

[0090] This structured latent response field mechanism achieves spatial continuity modeling from non-image index fluctuations to image region responses, avoiding the discrete mapping errors of the previous "point-to-point" approach. By combining Gaussian diffusion with spectral structure, the image enhancement is not only locally accurate but also globally structurally consistent. It has time-leading processing capabilities, which can identify the possible trends of early index fluctuations in future image changes. It can automatically adjust the diffusion range and index weights during the training process to adapt to different pathological scenarios.

[0091] To prioritize the expression of regions in structural images that indicate medical risks, highlighting risk-sensitive areas while preserving the integrity of the original visual semantics, and improving the ability of subsequent recognition modules to identify early signs of disease, this invention introduces a region regulation mechanism driven by the fluctuation of non-image indicators in the structural image processing stage. This mechanism uses the temporal fluctuation information of non-image indicators as a leading signal, and dynamically adjusts the expression intensity and visual weight of each region in the image through an established response mapping relationship with the image response region, thereby constructing an image expression space with indicator response-oriented characteristics. This regulation process not only enhances the expression clarity of regions highly correlated with indicators but also helps suppress redundant information in irrelevant regions, improving the overall image feature discrimination efficiency and sensitivity. Therefore, the image regulation module, based on the fluctuation of the patient's current non-image medical indicator data and in conjunction with the response mapping mechanism, performs expression enhancement or suppression operations on fluctuation-sensitive regions in the structural image. These operations adjust the image response intensity distribution of the target region to reflect its potential risk importance under the current indicator state.

[0092] Structural image information Represents the original retinal image in two-dimensional coordinates. The pixel grayscale value at that location. Non-image-related medical indicator data can be collected in the aforementioned data acquisition module and form a fluctuation feature vector under the current time window, denoted as:

[0093]

[0094] in, Indicates the first The standardized fluctuation values ​​of a non-graphical indicator (such as circulatory metabolism, blood pressure fluctuation, inflammatory factor concentration, etc.) within the current time window. This refers to the number of indicators.

[0095] Based on the fluctuation response mapping mechanism, the image control module calls the weight matrix:

[0096]

[0097] in, Indicates the first The first indicator for the first The control weights for each image region. Suppose the structural image is divided into... The image region set is a spatial region. Then the first The response adjustment factor for each region can be calculated as follows:

[0098]

[0099] in, For the region The adjustment factor is used to characterize the direction and magnitude of the impact of current index fluctuations on the image intensity of that region. When, it indicates that the regional expression should be enhanced, when When this is the case, it indicates that the expression of the region should be suppressed.

[0100] The system completes After calculation, for the region All pixel values ​​within Adjust the intensity as follows:

[0101]

[0102] in, This indicates the original image at the pixel level. grayscale value at that location This refers to the adjusted grayscale value of the image.

[0103] Preferably, to avoid edge artifacts caused by abrupt changes in image intensity between different regions, the system introduces a response transition band at the region boundaries and employs a spatial bilateral filtering algorithm to adjust the factor. Smoothing is performed to make the image response intensity change continuously in space, ensuring the naturalness and interpretability of the image.

[0104] In a preferred implementation, the image control module also includes a response filtering mechanism, that is, only when the regional response intensity... The control operation will only be performed when the following conditions are met:

[0105]

[0106] in, This is a preset control threshold used to prevent minor fluctuations from causing unnecessary image changes.

[0107] Through the collaborative processing of the fluctuation correlation module and the image modulation module, this system achieves deep fusion and dynamic linkage between non-image-related medical indicators and structural image information, effectively overcoming the limitation of traditional image feature extraction methods that cannot utilize the patient's historical physiological state. The fluctuation correlation module, by introducing a cross-modal response mapping mechanism, establishes a temporal correlation between non-image-related indicators and changes in image regions. This allows the system to identify sensitive areas with potential response trends before indicator fluctuations are significantly reflected in visually discernible image features. Based on this response relationship, the image modulation module dynamically adjusts the grayscale expression and structural clarity of local image regions, prioritizing the enhancement of tissue features related to high-risk indicator fluctuations and suppressing expression interference from low-correlation regions. This improves the discriminative power of image expression and the interpretability of disease states without altering the overall image structure. This combined mechanism not only enhances the early identifiability of key lesion regions in retinal images but also avoids interference from noise in irrelevant regions during feature extraction and risk assessment. It effectively enhances the model's predictive robustness and sensitivity under conditions of limited sample size or insignificant early pathology, providing a more clinically valuable image expression basis for the accurate prediction of the risk of retinal aneurysm rupture.

[0108] In a specific example, a 65-year-old male patient had a history of hypertension for over ten years and long-term metabolic abnormalities of high cholesterol and high triglycerides. During his annual fundus examination, the patient presented with no obvious symptoms and normal visual acuity. However, his blood pressure fluctuations had significantly increased over the past three months, and his C-reactive protein (CRP) level showed a sustained, slow upward trend. The data acquisition module collected retinal OCT and OCTA images on the day of the examination and simultaneously extracted time-series data of medical indicators such as blood pressure, CRP, low-density lipoprotein (LDL), and blood glucose. Through a fluctuation correlation module, the system established a cross-modal response mapping mechanism based on the joint distribution relationship between changes in blood pressure and CRP and changes in image microstructure in historical training samples. This mechanism identified a significant correlation between the current increase in CRP and the decrease in retinal small vessel wall density, particularly concentrated in the deep vascular structures of the nasal region of the optic disc.

[0109] Based on this, the image modulation module performs expression enhancement operations on the sensitive region in the original image. Specifically, the system sets a modulation factor for this region according to the fluctuations of the aforementioned indicators. The image underwent enhancement processing, including pixel grayscale value enhancement within the region and responsive smoothing at image edges to maintain natural image continuity. The enhanced image shows significantly clearer vessel wall boundaries and highlighted structural textures at vessel bifurcation points.

[0110] After performing expression enhancement or suppression operations on the structural image, in order to accurately characterize the local tissue state in the image and provide high-quality input features for subsequent risk prediction, it is necessary to perform deep encoding processing on the regulated image to extract discriminative image feature representations. This process should take into account both fine-grained texture variations and macroscopic structural layout in the image, and highlight the pathological expression differences in the regulated region, adapting to the sensitive recognition requirements of the subsequent identification module for risk patterns. Therefore, the image feature encoding module is used to encode the features of the structural image processed by the image regulation module, extracting image feature vectors reflecting the local tissue state.

[0111] The image feature encoding module is a functional module used to map structural image information into low-dimensional, structured vector representations. Essentially, it extracts the most expressive features from the image through multi-level feature extraction and nonlinear transformation operations to describe the potential pathological states of retinal tissue within the image. This module is typically implemented using convolutional neural networks, graph convolutional structures, or their derivatives to capture the changing features of local spatial structures in the image.

[0112] The image feature encoding module is designed based on the need for accurate modeling of local retinal tissue states in structural images. In the context of retinal aneurysm rupture risk prediction, and to adapt to the characteristics of weak early manifestations, atypical morphology, and minimal structural perturbation in lesion areas in images, this module aims to extract multi-scale feature representations reflecting pathological states from images by constructing an encoding path with discriminative power and spatial sensitivity. Therefore, the image feature encoding module is used to perform feature encoding on the structural images processed by the image modulation module, extracting image feature vectors reflecting local tissue states.

[0113] The image feature encoding module specifically includes an input preprocessing unit, a main encoding network structure, an attention modulation mechanism, and a feature fusion mapping structure. The input preprocessing unit performs normalized image cropping and pixel value normalization to meet the input standards of deep networks and improve the distribution consistency of image data from different sources. The main encoding network structure includes multiple concatenated convolutional and pooling layers. Convolutional layers extract local feature edges, pooling layers reduce spatial dimensions and enhance local invariance, and nonlinear activation functions nonlinearly enhance the linear mapping results. The convolution operation takes the form of:

[0114]

[0115] in, Indicates the first Layer output feature map, This indicates the input from the previous layer. For convolution kernel weights, For bias terms, It is a non-linear activation function.

[0116] To improve the encoding capability of high-response regions in an image, an attention modulation mechanism is preferably introduced after the convolutional layer. This mechanism includes two sub-modules: channel attention and spatial attention. The channel attention module measures the importance of each channel in the current modulation state, extracts global contextual information through global average pooling and max pooling, and constructs an attention weight vector. Then, a weighted fusion is performed with the original feature map. The spatial attention module is used to locate salient regions in the image and extract a two-dimensional attention map by combining the spatial gradient changes of the convolutional feature map. It is used to adjust the expressive intensity of each spatial location.

[0117] In a preferred implementation, the image feature encoding module employs a dual-branch structure and introduces residual connection paths. One branch encodes the overall structural layout and background configuration of the image, while the other branch focuses on high-response regions identified by the image modulation module, generating local and global feature maps respectively through convolutional pooling paths. The fusion structure uses a weighted fusion formula:

[0118]

[0119] in, and These represent the global and local encoding results, respectively. This is the fusion coefficient, which is dynamically adjusted based on the importance of the regulated area.

[0120] Subsequently, the fused feature maps are input to a fully connected layer or a global average pooling layer and mapped to low-dimensional image feature vectors. This vector serves as the input to the subsequent risk identification module, used to characterize the spatial distribution features of tissue state and lesion response characteristics in the current image.

[0121] Through the above structural design and coding process, the attention focus of the feature extraction path can be dynamically adjusted according to the response weight information provided by the image control module, so as to achieve accurate capture of early lesion areas. Through multi-scale and multi-branch structural design, the ability to jointly express image details and global patterns is enhanced. The introduced residual connection path can effectively alleviate the gradient vanishing problem in deep networks and improve the efficiency and stability of feature learning.

[0122] In this system, the risk identification module serves as the final judgment stage in the prediction process. Its core task is to perform pattern recognition on the image feature vectors output by the image feature encoding module to determine the rupture risk status of retinal aortic aneurysms. The rupture risk of retinal aortic aneurysms often stems from the combined effects of multiple physiological factors, such as potential structural changes in the vessel wall, abnormal tissue permeability, and local stress accumulation. These changes manifest as complex and subtle multidimensional patterns in the image features processed by the image modulation module and extracted by the feature encoding module, making them difficult to accurately classify through rule setting or expert experience.

[0123] By introducing a pattern recognition module with adaptive learning capabilities, a nonlinear mapping relationship between image feature vectors and risk levels can be established based on a large amount of historical sample data, enabling automatic identification of the risk status of retinal images. To this end, the risk identification module receives the image feature vectors and performs pattern recognition to determine the rupture risk status of retinal aneurysms.

[0124] The specific implementation of this module includes three steps: constructing a classification model, standardizing input features, and outputting recognition results. First, a classification model is constructed. Optional implementation schemes include using multi-layer neural network models, ensemble learning models, or graph neural network models. This model is trained on labeled samples to enable it to distinguish between high, medium, and low-risk images. Image feature vectors typically need to be normalized before being input into the classification model to eliminate feature shifts caused by differences in brightness, contrast, and scale between different images. During the model inference stage, the risk identification module, based on the learned parameters, classifies the input image feature vectors and outputs the corresponding risk level label or continuous risk score.

[0125] Preferably, in one preferred implementation, the risk identification module employs a parallel dual-channel structure. One channel extracts the global pattern of the overall image structure, while the other channel focuses on the feature information of high-response regions after modulation. The system dynamically adjusts the importance of the output features of the two channels through an attention mechanism and performs feature fusion in the final output stage to improve the sensitivity to discriminative changes in key regions. Furthermore, a dynamic adjustment mechanism for class weights can preferably be set to adapt to the problem of uneven distribution of high- and low-risk samples in different datasets, thereby improving the stability and generalization ability of the model under actual clinical data.

[0126] Through the aforementioned structural design, the risk identification module can automatically identify image feature patterns with a high tendency to break down, without relying on human experience intervention, and possesses good scalability and adaptability. This module can be used for dynamic tracking and analysis of images at continuous time points, and can also be combined with previously collected non-image data to construct a more comprehensive multidimensional risk assessment model, providing important auxiliary evidence for clinical decision-making.

[0127] For any module structures not specifically defined in this invention, the existing technical descriptions shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered part of this invention and used to understand the meaning of certain technical features or parameters.

Claims

1. A multi-index analysis system for predicting the risk of retinal aneurysm rupture, characterized in that, The system includes: The data acquisition module is used to collect patient-related structural image information and non-image medical indicator data. The structural image information is used to present the state of retinal vessels and surrounding tissues. The non-image medical indicator data includes time-series parameters that can characterize the patient's circulatory metabolic state and / or vascular stress changes and / or inflammatory response trends. The fluctuation correlation module is used to analyze the joint change relationship between the non-image-type medical indicator data and the corresponding structural image information in the time dimension based on historical sample data, and to establish a response mapping mechanism to reflect the fluctuation correlation pattern between the two types of data. The image modulation module is used to perform expression enhancement or expression suppression operations on regions in the structural image that are sensitive to fluctuations, based on the fluctuations of the current patient's non-image-related medical indicator data and in conjunction with the response mapping mechanism. The modulation operations are used to adjust the image response intensity distribution of the target region to reflect its potential risk importance under the current indicator state. The image feature encoding module is used to encode the features of the structural image after it has been processed by the image control module, and to extract image feature vectors that reflect the local tissue state. The risk identification module is used to receive the image feature vector and perform pattern recognition to determine the rupture risk status of the retinal aortic aneurysm.

2. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The non-image-based medical indicator data includes one or more of the following: blood pressure, heart rate variability, blood glucose level, blood lipid composition, and C-reactive protein level.

3. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The structural image information includes the ganglion cell layer, the nuclear layer, and the outer reticular layer.

4. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The analysis of the joint changes between the non-image-related medical indicator data and the corresponding structural image information over time, based on historical sample data, includes: Time standardization processing was performed on non-image-based medical indicator data; Extract local fluctuation features of non-image-related metrics within each sliding time window; Extract structural image information corresponding to the sliding time window; The structural image is divided into several regions based on anatomical structure or image segmentation algorithms; Extract preset image features from each divided region; Based on the Pearson correlation coefficient, the temporal correlation between different fluctuation feature dimensions and the features of each image region was calculated, and a statistical response map was established to reflect the degree of response of the fluctuation pattern to changes in image structure. Construct a heterogeneous graph structure, where the set of nodes includes fluctuating feature nodes and image response nodes, and edges represent potential response paths between the two. By using a graph neural network propagation mechanism, the connection weights between nodes are trained end-to-end, and an attention mechanism is introduced to assign the influence intensity of different indicators on image regions.

5. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The analysis of the joint changes between the non-image-related medical indicator data and the corresponding structural image information over time, based on historical sample data, includes: Construct a potential propagation response field based on the image region map, so that the fluctuation of each index diffuses on the image structure map in the form of Gaussian kernel weights, and dynamically forms a weight gradient distribution; For a non-image indicator in historical samples, the fluctuation characteristics are extracted within a sliding time window, and its dynamic fluctuation potential is extracted through a gating unit. The structural image is divided into a region map, where nodes represent local regions in the image and edges represent the spatial connections between regions. Construct a potential response weight field on the graph, where the response weight of each node is defined by a Gaussian diffusion function; The response fields of multiple indicators are synthesized into an overall response field according to a weighted average or maximization mechanism.

6. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The image modulation module specifically includes: Perform two-dimensional pixel grayscale conversion on structural image information; A weight matrix is ​​generated based on a fluctuation response mapping mechanism. The image of the corresponding region is processed and adjusted according to the weight matrix.

7. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 6, characterized in that, A response transition band is introduced at the region boundary, and a spatial bilateral filtering algorithm is used to smooth the adjustment factor, so that the image response intensity changes continuously in space.

8. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 6, characterized in that, Control operations will only be performed when the regional response intensity is greater than or less than the preset value.

9. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The image feature encoding module adopts a dual-branch structure and introduces residual connection paths. One branch encodes the overall structural layout and background configuration of the image, while the other branch focuses on the high-response regions identified by the image modulation module, generating local and global feature maps respectively through convolutional pooling paths.

10. The multi-index analysis system for predicting the risk of retinal aneurysm rupture according to claim 1, characterized in that, The risk identification module adopts a parallel dual-channel structure. One channel is used to extract the global pattern of the overall image structure, while the other channel focuses on the feature information of the high-response area after modulation. The system dynamically adjusts the importance of the output features of the two channels through an attention mechanism and performs feature fusion in the final output stage to improve the sensitivity to discriminative changes in key regions.

Images