Digital ophthalmology virtual intelligent doctor system based on artificial intelligence

Abstract

The invention relates to the technical field of medical image processing, in particular to a digital ophthalmology virtual intelligent doctor system based on artificial intelligence, and the system comprises an image obtaining module which is used for obtaining and preprocessing a fundus image to form a standardized image; the blood vessel segmentation module is used for extracting a blood vessel network, constructing a blood vessel topological map and recording blood vessel branch and path information; the micro-hemorrhage candidate generation module is used for screening red spots and performing path matching in combination with vascular topology to form candidate spots; the staged reasoning module is used for roughly scanning and finely classifying the candidate spots to generate micro bleeding points; the historical image management module is used for comparing the candidate points with historical images of the patient, marking new or changing spots and forming priority information; and the output and feedback module is used for generating a detection result according to the candidate micro bleeding points and the priority information. Automatic identification and sorting of the micro bleeding points are realized, and the screening efficiency and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a digital ophthalmology virtual intelligent doctor system based on artificial intelligence. Background Technology

[0002] Detection of fundus microbleeds mainly relies on manual image diagnosis or analysis using traditional image processing algorithms. Manual diagnosis depends on the experience of radiologists to identify spots, which is time-consuming and easily affected by subjective factors. Traditional image processing methods typically screen red spots through color thresholding, morphological analysis, or texture feature extraction, and combine some rules to judge candidate microbleeds, which can achieve automatic detection to a certain extent. Some methods also attempt to use deep learning models for global classification or pixel-level segmentation to improve the level of automation in detection. These methods can identify some lesion areas under ideal image conditions, but still require manual verification.

[0003] Existing technologies are prone to false positives and false negatives when dealing with densely packed or intersecting vascular areas. Microbleeds are small in size, have diverse shapes, and are similar in color to blood vessels and other red areas. They are also subject to severe interference from background textures, making it difficult for traditional methods to accurately distinguish between real lesions and vascular structures, and thus failing to achieve highly reliable automatic identification of microbleeds. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a digital ophthalmology virtual intelligent doctor system based on artificial intelligence, which aims to improve the problem that traditional methods are unable to accurately distinguish between real lesions and vascular structures, and cannot achieve highly reliable automatic identification of microbleeding points.

[0005] In a first aspect, the present invention provides the following technical solution: a digital ophthalmology virtual intelligent doctor system based on artificial intelligence, comprising the following modules: The image acquisition module is used to acquire images of the patient's fundus and preprocess the images to form preprocessed images; The blood vessel segmentation module is used to segment blood vessels in preprocessed images, extract blood vessel networks, construct blood vessel topology maps, and record blood vessel branch, path, and direction information. The microbleed candidate generation module is used to initially screen red spots in the preprocessed image and perform path matching in combination with the vascular topology map to form candidate spot information; The phased reasoning module is used to perform coarse scanning and fine classification of candidate spot information to form candidate microbleed points; The historical image management module is used to combine the patient's historical fundus images and vascular topology information, compare candidate microbleeds with historical candidate points, mark new or changed spots, and form candidate region priority information. The output and feedback module is used to generate microbleed detection results based on candidate microbleeds and candidate region priority information.

[0006] By adopting the above technical solution, the red spots in fundus images are constrained and screened by combining the vascular topology, which can automatically distinguish microbleeds from blood vessels and other red areas and accurately extract candidate microbleeds. This solves the problem of false detection and false negative detection in existing technologies in areas with dense or intersecting blood vessels, and achieves highly reliable automatic microbleed identification.

[0007] Preferably, the image preprocessing includes: Acquire raw fundus images and standardize the image data format and spatial coordinates; The standardized image is subjected to grayscale normalization to form a uniform grayscale distribution range; The normalized image is subjected to noise suppression processing, and random noise in the image is smoothed. Tissue boundary enhancement processing is performed on the noise-suppressed image to improve the distinction between vascular regions and background regions; A preprocessed image is constructed based on the image with enhanced boundaries.

[0008] Preferably, the blood vessel segmentation includes: Receive the preprocessed image and determine the target fundus region in the image; Candidate blood vessel region extraction processing is performed on the target fundus region to obtain an initial region where blood vessels may be distributed; Local structural analysis is performed on the initial region to generate characteristic data representing the tubular structure of blood vessels; Based on the feature data, the vascular region and non-vascular region in the image are distinguished to obtain the vascular region segmentation result.

[0009] Preferably, the construction of the vascular topology map is as follows: Receive the segmented blood vessel image obtained by blood vessel segmentation, and perform skeletonization processing on the segmented blood vessel image to obtain the blood vessel centerline structure; Node identification is performed on the central line structure to determine the bifurcation points, termination points, and main trunk nodes of the blood vessels; The path of the centerline segment between any two adjacent nodes is marked to form path data representing the direction of blood vessels; Based on the path data, the branching relationships, connection relationships, and path lengths of blood vessels are recorded to construct a blood vessel topology. The vascular topology is mapped to form a vascular topology atlas.

[0010] Preferably, the screening of red spots: Receive the preprocessed image obtained through preprocessing, and perform color space conversion on the image to obtain a target color space image containing red channel information; A red response threshold is set for the target color space image, and the pixel regions that meet the threshold condition are initially marked. Perform connected component analysis on the initial region to extract the connected regions of suspected red spots; For each connected region, calculate its area, roundness, and grayscale uniformity indices, screen out regions that do not meet the morphological characteristics of microbleeds, and generate a preliminary set of red spots.

[0011] Preferably, the path matching includes: Receive candidate spot data generated by the red spot filtering step, and read the vessel branch node and path coordinate information from the vessel topology map; Based on the location coordinates of the candidate spots, the adjacent vascular branches or path segments are determined as matching reference areas. For each candidate spot, calculate its spatial distance to the matching reference region, and filter the spots that can participate in path association according to the preset location association criteria; Perform directional consistency analysis on the selected spots to determine the angular relationship between the spot center and the direction of the blood vessel path, and filter out spots that do not meet the conditions according to the preset angle range; Spots that meet the location association criteria and orientation consistency conditions are marked, and candidate spot information after path matching is output.

[0012] Preferably, the coarse scan includes: Receive candidate blob information after path matching processing, and extract local image patches of each blob as coarse inference input; Based on a preset lightweight feature extraction network, low-dimensional feature extraction is performed on the local image patch to obtain color distribution, edge structure and local texture features; The extracted low-dimensional features are quickly classified using a coarse classification model to obtain a rough risk label for each candidate spot. Based on the rough risk labels, a set of spots that need to enter the fine classification stage is selected, and the corresponding candidate region coordinate information is recorded.

[0013] Preferably, the fine classification includes: Receive the set of candidate spots obtained by coarse scanning and screening, and perform high-resolution resampling processing on the image region corresponding to each candidate spot; Based on a high-resolution feature extraction network, multi-scale texture features, structural features, and chromaticity detail features are extracted from the resampled image region. The multi-scale features are input into a deep classification model to obtain a deep classification feature vector for each candidate blob. The classification decision module is used to determine the category of the deep classification feature vector and generate the corresponding microbleed point classification result. The classification results are labeled with their locations, and corresponding candidate microbleed point information is generated.

[0014] Preferably, comparing candidate microbleed points with historical candidate points includes: Receive candidate microbleed point information obtained through fine classification in the current image; Retriev historical fundus images and corresponding historical candidate microbleeds from the historical image database; The current candidate microbleed points are matched with historical candidate points in terms of spatial location to identify newly emerging candidate points that overlap, are adjacent, or have not yet appeared. Time-series labeling was performed on the matching results to distinguish between stable, regressing, or newly emerging candidate microbleeds.

[0015] Preferably, the generated microbleed point detection results include: Receive a set of candidate microbleeds obtained by comparing historical images and their time-series annotation information; Based on the location, classification results, and time series annotation information of the candidate microbleeds, the priority weight of each candidate microbleed is calculated. The priority weights are integrated with the candidate microbleed point location coordinates and classification labels to form a microbleed point detection dataset; Microbleed detection results are generated based on the microbleed detection dataset, including the spatial location and corresponding priority information of candidate microbleeds.

[0016] The present invention has the following beneficial effects: 1. In this invention, by systematically segmenting blood vessels in the patient's fundus image and further constructing a vascular topology map, a reasonable matching and screening of candidate microbleeding points and vascular paths is achieved. Through comprehensive analysis of vascular branch, path and direction information, true microbleeding points can be accurately identified in dense or intersecting vascular areas, while red spots that do not conform to the vascular structure are removed. This solves the problem of false detection and false negative detection caused by the small size, diverse shape and color of microbleeding points that are close to blood vessels or other red areas and are easily affected by background texture interference, thus improving the reliability and accuracy of the detection results.

[0017] 2. In this invention, the artificial intelligence detection model is divided into a coarse scanning stage and a fine classification stage. In the coarse stage, a low-resolution rapid scan is used to generate candidate regions. In the fine stage, high-resolution deep learning features are extracted and classified from the candidate regions. At the same time, by combining lightweight model design, quantization and pruning techniques, efficient operation on low-computing-power mobile devices is achieved. This solves the problem that grassroots hospitals or mobile screening devices have limited computing power and cannot directly run high-complexity models to complete fine microbleed detection, and ensures a balance between detection speed and accuracy.

[0018] 3. In this invention, by collecting historical fundus images of patients and establishing a database of candidate microbleeds, the spatial location and time sequence of candidate spots in the current image are compared with those in the historical image. This enables the identification of newly formed or changing microbleeds and assigns them higher priority. At the same time, the inference frequency of stable spots is reduced. This solves the problems that a single fundus image analysis cannot distinguish between newly formed and stable microbleeds, that low-computing-power devices may repeatedly calculate stable spots, resulting in wasted inference resources and low screening efficiency. It achieves dynamic priority ranking and optimization of the detection process. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of a digital ophthalmology virtual intelligent doctor system based on artificial intelligence proposed in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: In a first embodiment of the present invention, the present invention provides a digital ophthalmology virtual intelligent doctor system based on artificial intelligence, such as... Figure 1 As shown, it includes the following modules: The image acquisition module is used to acquire images of the patient's fundus and preprocess the images to form preprocessed images; Further image preprocessing includes: Acquire raw fundus images and standardize the image data format and spatial coordinates; The standardized image is subjected to grayscale normalization to form a uniform grayscale distribution range; The normalized image is subjected to noise suppression processing, and random noise in the image is smoothed. Tissue boundary enhancement processing is performed on the noise-suppressed image to improve the distinction between vascular regions and background regions; A preprocessed image is constructed based on the image with enhanced boundaries.

[0022] Specifically, the patient's fundus images are first acquired through an image acquisition module. This acquisition can be done using a conventional fundus camera or digital imaging equipment. The acquired raw images have their original resolution, color channels, and spatial coordinate information. Subsequently, the data format of the raw images is standardized to conform to the system's processing specifications, including the standardization of image pixel arrangement format, file encoding format, and image size. This ensures that fundus images from different sources can maintain consistency in subsequent processing. At the same time, the spatial coordinates of the images are standardized so that the spatial position of each pixel in the image corresponds to the actual fundus region, facilitating subsequent vascular localization and microbleed point identification. After image standardization, grayscale normalization is performed, adjusting the grayscale value of each pixel in the image. Perform mapping to make it fall within the preset grayscale range. Internally, grayscale normalization can be achieved using a linear transformation method: in, Indicates the pixel position of the original image. grayscale value, and These represent the minimum and maximum gray values ​​of the original image, respectively. This represents the normalized grayscale value. and These are the lower and upper limits of the normalized grayscale range, respectively. This processing makes images acquired under different lighting conditions and with different devices comparable, providing a unified grayscale basis for subsequent extraction of blood vessels and red spots. After grayscale normalization, noise suppression processing is performed on the image to smooth out random noise or high-frequency interference introduced during the acquisition process. Noise suppression can be achieved using a Gaussian filtering algorithm; for example, when using Gaussian filtering, the image grayscale value update formula is: in, This indicates the position of the normalized image. grayscale value, This represents the smoothed grayscale value after filtering. The Gaussian kernel function is defined as follows: in, These are the offset coordinates within the filtering window. The standard deviation of the Gaussian kernel is used to control the degree of smoothing. The radius of the filtering window is used; this processing can reduce image noise while preserving the blood vessel edge structure, providing a clearer image basis for blood vessel segmentation and microbleed detection. After noise suppression, the image undergoes further tissue boundary enhancement processing to improve the distinguishability between vascular regions and background regions. Boundary enhancement can be combined with gradient calculation and local contrast enhancement, such as calculating the image gradient magnitude. in, and These represent the grayscale changes in the horizontal and vertical directions, respectively. The gradient magnitude at this pixel location is used to characterize the blood vessel boundary; subsequently, the gradient magnitude is weighted and fused with the smoothed image to generate the boundary-enhanced image. in, and These are the weighting coefficients for the image and gradient, respectively, used to balance brightness and boundary information; Preprocessed image based on boundary enhancement image The image contains standardized grayscale, noise suppression, and vascular boundary enhancement information, which can be directly used as input for the vascular segmentation module and the microbleed candidate generation module. This preprocessed image ensures the consistency of image data format and spatial coordinates, and provides a visual feature foundation suitable for deep learning networks or traditional image processing algorithms, enabling the entire digital ophthalmology virtual intelligent doctor system to continuously and stably analyze and process patient fundus images.

[0023] The blood vessel segmentation module is used to segment blood vessels in preprocessed images, extract blood vessel networks, construct blood vessel topology maps, and record blood vessel branch, path, and direction information. Further, vascular segmentation includes: Receive the preprocessed image and determine the target fundus region in the image; Candidate vessel region extraction is performed on the target fundus region to obtain the initial region where vessels may be distributed; Local structural analysis is performed on the initial region to generate characteristic data representing the tubular structure of blood vessels; Based on feature data, the vascular region and non-vascular region in the image are distinguished to obtain the vascular region segmentation result.

[0024] Furthermore, construct a vascular topology atlas: Receive the segmented blood vessel image obtained by blood vessel segmentation, and perform skeletonization processing on the segmented blood vessel image to obtain the blood vessel centerline structure; Node identification is performed on the centerline structure to determine the bifurcation points, termination points, and main trunk nodes of the blood vessels; The path of the centerline segment between any two adjacent nodes is marked to form path data representing the direction of blood vessels; Based on path data, the branching relationships, connection relationships, and path lengths of blood vessels are recorded to construct the blood vessel topology. The topological structure of blood vessels is mapped to form a vascular topology atlas.

[0025] Specifically, the blood vessel segmentation module first receives a preprocessed image, which has undergone grayscale normalization, noise suppression, and boundary enhancement, providing clear information about the blood vessel structure. The system identifies the target fundus region in the image, prioritizing areas with a high probability of blood vessel distribution. The target fundus region can be determined through morphological feature analysis, brightness thresholding, or deep learning localization networks. After determining the region, candidate blood vessel region extraction is performed. By analyzing the local grayscale, edge gradient, and texture features of the image, an initial blood vessel region mask is generated. This mask marks the set of pixels that may contain blood vessels, providing a foundation for subsequent fine segmentation. After obtaining the initial region, local structural analysis is performed on each candidate vessel region to extract vessel tubular structure features, such as local width, orientation, curvature, and neighborhood connectivity. These feature data can be represented in vector form. ,in Indicates the local width of the blood vessel. Indicates the local direction angle. Indicates curvature. This represents a pixel connectivity index; by analyzing these feature data, candidate regions can be classified, distinguishing real blood vessel regions from non-vascular interference regions, thus obtaining accurate blood vessel segmentation results. Each pixel Indicates blood vessels. This indicates the background; the segmentation result provides direct input for subsequent vascular topology mapping while ensuring the integrity and continuity of the vascular structure. After vessel segmentation, a vessel topology map is constructed. The segmented vessel image is then skeletonized to extract the vessel centerline structure. Skeletonization can be achieved through distance transformation or thinning algorithms, reducing the vessel width information to a single-pixel centerline. Node recognition analysis of the skeleton image identifies vessel bifurcation points, termination points, and main trunk nodes. Each node can be represented as... ,in For node coordinates, Node types are branched, terminated, or trunked; the centerline segment between any two adjacent nodes generates path data through path marking. , indicating from node To the node A continuous set of pixel coordinates is used to represent the direction of blood vessels; the branching relationships, connection relationships, and path length of blood vessels are recorded based on path data. This forms the vascular topology, in which Indicates the path length; The topological structure is mapped to generate a vascular topological atlas. ,in For a set of nodes, For a set of paths, This map provides path length information; it also provides spatial connectivity, branching structure, and orientation information of the vascular network, which can be directly used for path matching of microbleed candidate points and comparison of historical images. This ensures that the system can accurately associate the vascular direction with the lesion location during microbleed detection and fundus analysis, and achieve continuous and structured vascular information processing.

[0026] The microbleed candidate generation module is used to initially screen red spots in the preprocessed image and perform path matching in combination with the vascular topology map to form candidate spot information; Further, filter out the red spots: Receive the preprocessed image obtained through preprocessing, and perform color space conversion on the image to obtain a target color space image containing red channel information; A red response threshold is set for the target color space image, and the pixel regions that meet the threshold condition are initially marked. Perform connected component analysis on the initial region to extract the connected regions suspected of being red spots; For each connected region, calculate its area, roundness, and grayscale uniformity indices, screen out regions that do not meet the morphological characteristics of microbleeds, and generate a preliminary set of red spots.

[0027] Furthermore, path matching includes: Receive candidate spot data generated by the red spot filtering step, and read the vessel branch node and path coordinate information from the vessel topology map; Based on the location coordinates of the candidate spots, the adjacent vascular branches or path segments are determined as matching reference areas. For each candidate spot, calculate its spatial distance to the matching reference area, and filter the spots that can participate in path association according to the preset location association criteria; Perform directional consistency analysis on the selected spots to determine the angular relationship between the spot center and the direction of the blood vessel path, and filter out spots that do not meet the conditions according to the preset angle range; Spots that meet the location association criteria and orientation consistency conditions are marked, and candidate spot information after path matching is output.

[0028] Specifically, the microbleed candidate generation module first receives a preprocessed image, which undergoes grayscale normalization, noise suppression, and boundary enhancement to provide clear information on fundus blood vessels and background structures. The system then converts the image from RGB space to a target color space that includes red channel information, for example, converting an RGB image to the red channel in HSV or Lab space, thus generating a red channel image. ,in Represents pixel coordinates, Indicates the red response value of the pixel; In the red channel image, set a red response threshold. , for satisfying The pixel regions are initially marked to generate a set of red candidate pixels. These candidate pixels are used for the next step of connected component analysis; connected component analysis is performed on the initially labeled region to detect the spatial connectivity of the red candidate pixels and extract the connected regions. Each region contains its set of pixel coordinates. Then, the area of ​​each connected region is calculated. Roundness and grayscale uniformity index ,in Let the perimeter of the connected region be . The average red response value for the region; these indicators are used to screen out areas that do not meet the morphological characteristics of microbleeds, forming a preliminary set of red spots. ; After the initial screening of red spots, the path matching step combines the vascular topology map to verify and associate candidate spots; the system reads the vascular branch nodes and path coordinate information from the vascular topology map. ,in For a set of nodes, For a set of paths, For path length information; for each candidate red spot According to its coordinates Identify adjacent vascular branches or pathway segments as matching reference regions; calculate the shortest spatial distance from the spot to the reference path. ,in Given the pixel coordinates on the path segment, candidate spots that meet the conditions are selected based on the preset position association criteria; Further directional consistency analysis is performed on the spots that pass the location screening to calculate the directional angle between the spot center and the corresponding path segment. ,in The vector representing the relative path between the center of the spot. This represents the path direction vector, with the slope or angle falling within a preset range. The blobs within are marked as conforming to orientation consistency; ultimately, a set of blobs that satisfy both the positional association criterion and the orientation consistency condition will be defined. The output forms candidate spot information, providing input for the subsequent phased reasoning module to perform coarse scanning and fine classification. This process ensures that microbleed candidate points consider both color response features and spatial relationships of vascular structures, achieving structured and location-related candidate generation.

[0029] The phased reasoning module is used to perform coarse scanning and fine classification of candidate spot information to form candidate microbleed points; Further, coarse scans include: Receive candidate blob information after path matching processing, and extract local image patches of each blob as coarse inference input; Based on a pre-defined lightweight feature extraction network, low-dimensional features are extracted from local image patches to obtain color distribution, edge structure, and local texture features. The extracted low-dimensional features are quickly classified using a coarse classification model to obtain a rough risk label for each candidate spot. Based on the rough risk labels, select the set of spots that need to enter the fine classification stage, and record the corresponding candidate region coordinate information.

[0030] Further, more detailed classifications include: Receive the set of candidate spots obtained by coarse scanning and screening, and perform high-resolution resampling processing on the image region corresponding to each candidate spot; Based on a high-resolution feature extraction network, multi-scale texture features, structural features, and chromaticity detail features are extracted from the resampled image region. Multi-scale features are input into a deep classification model to obtain a deep classification feature vector for each candidate blob. The classification decision module is used to determine the category of the deep classification feature vector and generate the corresponding microbleed point classification results. The classification results are labeled with their locations, and corresponding candidate microbleed point information is generated.

[0031] Specifically, the phased inference module first receives candidate blob information generated by the path matching step. The coordinates of each candidate blob and the path information of nearby blood vessels are used to determine the extraction range of the local image patch. The system then extracts a local image patch centered on the candidate blob from the preprocessed image. ,in The pixel coordinates are used, and the size of the local image patch can be set based on the average size of the microbleed points, for example, as... Pixels, to ensure that the spots and their surrounding blood vessels and background structures are included; During the coarse scanning stage, local image patches are input into a pre-defined lightweight feature extraction network, such as a convolutional neural network or a lightweight residual network, to extract low-dimensional feature vectors from the image patches. ,in The feature dimensions are represented; these feature vectors characterize color distribution, edge structure, and local texture patterns, such as color histograms, gradient orientation histograms, and local binary patterns; subsequently, the coarse classification model is based on... Perform rapid classification on each candidate spot and output a coarse risk label. Where 0 represents non-microbleeds and 1 represents suspected microbleeds; the system uses coarse risk labels to filter out a set of candidate spots that need to enter the fine classification stage. And record the local image coordinates corresponding to each spot. For precise classification and use; In the fine classification stage, the candidate blob set Each spot in the image is subjected to high-resolution resampling of its corresponding region to generate an enhanced image patch. To fully preserve minute texture and chromaticity information, the resampled image patches are input into a high-resolution feature extraction network, such as a multi-scale convolutional neural network or a deep feature pyramid network, to extract multi-scale texture features from the image patches. Structural features and color detail features These features combine to form a deep classification feature vector. The vector length is related to the multi-scale feature dimension; Deep classification feature vector Input a deep classification model, such as a deep convolutional network or a Transformer structure, to obtain the deep classification feature output for each candidate blob. , representing the probability of microbleeds; the classification decision module for Perform category determination, for example, by threshold. If microbleeds are present, If a spot is marked as a microbleed, it is marked as a non-microbleed; the system also labels the spatial location of each spot while generating the classification results, and outputs candidate microbleed information. ,in This represents the classification label for microbleeds. Through this process, the phased reasoning module achieves a complete operation from rapid screening of candidate spots to in-depth classification verification, ensuring that the microbleed information is structured, localizable, and provides complete input for subsequent historical image comparison and detection result generation.

[0032] The historical image management module is used to combine the patient's historical fundus images and vascular topology information, compare candidate microbleeds with historical candidate points, mark new or changed spots, and form candidate region priority information. Furthermore, comparing candidate microbleed points with historical candidate points includes: Receive candidate microbleed point information obtained through fine classification in the current image; Retriev historical fundus images and corresponding historical candidate microbleeds from the historical image database; The current candidate microbleed points are matched with historical candidate points in terms of spatial location to identify newly emerging candidate points that overlap, are adjacent, or have not yet appeared. Time-series labeling was performed on the matching results to distinguish between stable, regressing, or newly emerging candidate microbleeds.

[0033] Specifically, the historical image management module first receives candidate microbleed point information obtained through fine classification in the current image. Each candidate microbleed point contains its spatial coordinates. and category tags The system uses this information to retrieve the patient's historical fundus images and corresponding historical candidate microbleeds from the historical image database. Each historical candidate also includes its spatial coordinates. and time tags ; The system spatially matches the current candidate microbleed points with historical candidate points. The matching algorithm can be based on Euclidean distance. ,in Indicates the current candidate point Compared with historical candidate points Spatial distance, if If they are located in the same or adjacent positions, then they are determined to be either overlapping or adjacent. The preset spatial threshold can be set according to the resolution of the fundus image and the average size of the microbleeds; for current spots that do not match historical candidate points, they are marked as new candidate points for further tracking. After completing spatial matching, the system annotates the matching results with time-series information, distinguishing between stable microbleeds, faded microbleeds, and newly formed microbleeds. Stable microbleeds are those that have been matched in multiple consecutive historical images; faded microbleeds are those that exist in historical records but have not been matched in the current image; and newly formed microbleeds are those that are not matched in historical images for the current candidate microbleed. The system further generates candidate region priority information based on the matching results and time-series annotations. Priority can be calculated by weighting factors such as neoplasia, persistence, and the importance of adjacent blood vessels, for example... ,in Indicates stability index, Indicators of neonatal sex Indicators representing vascular proximity, To ensure that candidate microbleeds are properly sorted in the subsequent detection results generation, corresponding weight parameters are used. Through the above process, the historical image management module achieves a comprehensive comparison between the current microbleed candidate points and historical image data, time series annotation, and priority generation of candidate regions, providing structured and traceable microbleed information for the subsequent output and feedback modules. This ensures that the entire digital ophthalmology virtual intelligent doctor system can continuously track changes in the patient's fundus and form operable data support.

[0034] The output and feedback module is used to generate microbleed detection results based on candidate microbleeds and candidate region priority information.

[0035] Furthermore, the generated microbleed point detection results include: Receive a set of candidate microbleeds obtained by comparing historical images and their time-series annotation information; Based on the location, classification results, and time series annotation information of the candidate microbleeds, the priority weight of each candidate microbleed is calculated. The priority weights are integrated with the candidate microbleed point location coordinates and classification labels to form a microbleed point detection dataset; The microbleed detection results are generated based on the microbleed detection dataset, including the spatial location and corresponding priority information of candidate microbleeds.

[0036] Specifically, the output and feedback module first receives a set of candidate microbleed points obtained through historical image comparison, and each candidate microbleed point contains its spatial coordinates. Category tags and time series annotation information The module calculates the priority weight of each candidate microbleed point based on this information. The priority weights can comprehensively consider the neonatality, stability, importance of adjacent blood vessels, and classification confidence of microbleeding points. It can be calculated using a weighted formula: in The stability index of a time series is the ratio of the number of consecutive occurrences to the total number of historical observations. This indicates the novelty index, which is 1 if the current point has not appeared in the historical record, and 0 otherwise. This is an indicator of vascular proximity, calculated based on the shortest distance between microbleeding points and the main trunk or branch nodes of blood vessels. Represents the classification confidence level, which is the probability of the classification result based on the deep features output during the fine-grained classification stage; These are preset weighting coefficients used to balance the contributions of different factors to the priority level. After completing the priority calculation, the module assigns the spatial coordinates of each candidate microbleed point. Category tags and the calculated priority weights The data will be integrated to form a microbleeding point detection dataset. The dataset structure may include candidate point identifiers, spatial locations, classification labels, time series status, and priority values, providing data support for subsequent visualization or clinical decision-making. Based on the microbleeding point detection dataset, the module generates microbleeding point detection results. The spatial location and corresponding priority information of each candidate microbleeding point can be output through graphical annotation or data interface, which makes it easier for doctors or intelligent systems to focus on key areas and conduct further analysis. The generation process can be combined with threshold filtering or sorting algorithms to highlight high-priority points, and can also update the patient's historical records in a synchronous manner so as to continue to track the changing trend of microbleeding points in future image analysis. Through the above process, the output and feedback module realizes the structured integration, priority sorting and detection result generation of candidate microbleeding point information, ensuring that the system can completely feed back the spatial location, time series status and risk level information of microbleeding points to the clinical application end or subsequent processing module, providing operable and traceable microbleeding point detection results for the entire digital ophthalmology virtual intelligent doctor system.

[0037] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A digital ophthalmology virtual intelligent doctor system based on artificial intelligence, characterized in that, Includes the following modules: The image acquisition module is used to acquire images of the patient's fundus and preprocess the images to form preprocessed images; The blood vessel segmentation module is used to segment blood vessels in preprocessed images, extract blood vessel networks, construct blood vessel topology maps, and record blood vessel branch, path, and direction information. The microbleed candidate generation module is used to initially screen red spots in the preprocessed image and perform path matching in combination with the vascular topology map to form candidate spot information; The phased reasoning module is used to perform coarse scanning and fine classification of candidate spot information to form candidate microbleed points; The historical image management module is used to combine the patient's historical fundus images and vascular topology information, compare candidate microbleeds with historical candidate points, mark new or changed spots, and form candidate region priority information. The output and feedback module is used to generate microbleed detection results based on candidate microbleeds and candidate region priority information.

2. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The image preprocessing includes: Acquire raw fundus images and standardize the image data format and spatial coordinates; The standardized image is subjected to grayscale normalization to form a uniform grayscale distribution range; The normalized image is subjected to noise suppression processing, and random noise in the image is smoothed. Tissue boundary enhancement processing is performed on the noise-suppressed image to improve the distinction between vascular regions and background regions; A preprocessed image is constructed based on the image with enhanced boundaries.

3. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The blood vessel segmentation includes: Receive the preprocessed image and determine the target fundus region in the image; Candidate blood vessel region extraction processing is performed on the target fundus region to obtain an initial region where blood vessels may be distributed; Local structural analysis is performed on the initial region to generate characteristic data representing the tubular structure of blood vessels; Based on the feature data, the vascular region and non-vascular region in the image are distinguished to obtain the vascular region segmentation result.

4. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The construction of the vascular topology map: Receive a segmented image of blood vessels obtained by blood vessel segmentation, and perform skeletonization processing on the segmented image of blood vessels to obtain the centerline structure of blood vessels; Node identification is performed on the central line structure to determine the bifurcation points, termination points, and main trunk nodes of the blood vessels; The path of the centerline segment between any two adjacent nodes is marked to form path data representing the direction of blood vessels; Based on the path data, the branching relationships, connection relationships, and path lengths of blood vessels are recorded, and a blood vessel topology is constructed. The vascular topology is mapped to form a vascular topology atlas.

5. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The screening of red spots: Receive the preprocessed image obtained through preprocessing, and perform color space conversion on the image to obtain a target color space image containing red channel information; A red response threshold is set for the target color space image, and the pixel regions that meet the threshold condition are initially marked. Perform connected component analysis on the initial region to extract the connected regions of suspected red spots; For each connected region, calculate its area, roundness, and grayscale uniformity indices, screen out regions that do not meet the morphological characteristics of microbleeds, and generate a preliminary set of red spots.

6. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The path matching includes: Receive candidate spot data generated by the red spot filtering step, and read the vessel branch node and path coordinate information from the vessel topology map; Based on the location coordinates of the candidate spots, the adjacent vascular branches or path segments are determined as matching reference areas. For each candidate spot, calculate its spatial distance to the matching reference region, and filter the spots that can participate in path association according to the preset location association criteria; Perform directional consistency analysis on the selected spots to determine the angular relationship between the spot center and the direction of the blood vessel path, and filter out spots that do not meet the conditions according to the preset angle range; Spots that meet the location association criteria and orientation consistency conditions are marked, and candidate spot information after path matching is output.

7. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The coarse scan includes: Receive candidate blob information after path matching processing, and extract local image patches of each blob as coarse inference input; Based on a preset lightweight feature extraction network, low-dimensional feature extraction is performed on the local image patch to obtain color distribution, edge structure and local texture features; The extracted low-dimensional features are quickly classified using a coarse classification model to obtain a rough risk label for each candidate spot. Based on the rough risk labels, a set of spots that need to enter the fine classification stage is selected, and the corresponding candidate region coordinate information is recorded.

8. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The detailed classification includes: Receive the set of candidate spots obtained by coarse scanning and screening, and perform high-resolution resampling processing on the image region corresponding to each candidate spot; Based on a high-resolution feature extraction network, multi-scale texture features, structural features, and chromaticity detail features are extracted from the resampled image region. The multi-scale features are input into a deep classification model to obtain a deep classification feature vector for each candidate blob. The classification decision module is used to determine the category of the deep classification feature vector and generate the corresponding microbleed point classification result. The classification results are labeled with their locations, and corresponding candidate microbleed point information is generated.

9. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The comparison of candidate microbleed points with historical candidate points includes: Receive candidate microbleed point information obtained through fine classification in the current image; Retriev historical fundus images and corresponding historical candidate microbleeds from the historical image database; The current candidate microbleed points are matched with historical candidate points in terms of spatial location to identify newly emerging candidate points that overlap, are adjacent, or have not yet appeared. Time-series labeling was performed on the matching results to distinguish between stable, regressing, or newly emerging candidate microbleeds.

10. The digital ophthalmology virtual intelligent doctor system based on artificial intelligence according to claim 1, characterized in that, The generated microbleed point detection results include: Receive a set of candidate microbleeds obtained by comparing historical images and their time-series annotation information; Based on the location, classification results, and time series annotation information of the candidate microbleeds, the priority weight of each candidate microbleed is calculated. The priority weights are integrated with the candidate microbleed point location coordinates and classification labels to form a microbleed point detection dataset; Microbleed detection results are generated based on the microbleed detection dataset, including the spatial location and corresponding priority information of candidate microbleeds.

Images