Search

CN-121982008-A - Grading diagnosis system and method for diabetic retinopathy

CN121982008ACN 121982008 ACN121982008 ACN 121982008ACN-121982008-A

Abstract

The invention discloses a grading diagnosis system and method for diabetic retinopathy, belonging to the technical field of medical image processing and computer vision, wherein the system comprises a multi-focus cooperative detection module, a vascular topological graph construction module, a hierarchical grading decision module and a self-adaptive feedback optimization module, wherein the multi-focus cooperative detection module recognizes five types of typical focuses and calculates focus cooperative perception indexes; the vessel topological graph construction module learns vessel topological characteristics through a graph neural network; the self-adaptive feedback optimization module dynamically adjusts detection parameters according to the confidence score to form a closed-loop optimization mechanism, the invention fully utilizes focus cooperative information and a vascular topological structure, realizes accurate and reliable diabetic retinopathy hierarchical diagnosis through a deep coupling closed-loop cooperative system, and is suitable for fundus screening of basic medical institutions.

Inventors

  • Tai Yuqing
  • GE TIANQI

Assignees

  • 山东中医药大学

Dates

Publication Date
20260505
Application Date
20260129

Claims (10)

  1. 1. A hierarchical diagnosis system for diabetic retinopathy, comprising: The multi-lesion cooperative detection module is used for carrying out feature extraction on the eye bottom image based on the shared feature encoder, identifying five types of lesions of microaneurysms, hard exudation, cotton-wool spots, bleeding foci and new blood vessels, outputting a spatial distribution thermodynamic diagram and a counting statistic of each type of lesions, and calculating a lesion cooperative perception index based on the spatial proximity and the co-occurrence mode among the lesions; the blood vessel topological graph construction module is connected with the multi-lesion collaborative detection module and is used for carrying out blood vessel segmentation on the fundus image, extracting a blood vessel center line and branch nodes, constructing a blood vessel topological graph, taking the blood vessel nodes as graph nodes, taking the blood vessel segments as graph edges, carrying out embedded learning on the blood vessel topological graph through a graph neural network, and generating a blood vessel topological feature vector; the hierarchical grading decision module is connected with the multi-focus collaborative detection module and the vascular topological graph construction module and is used for carrying out multi-scale feature fusion on the focus collaborative perception index, the spatial distribution thermodynamic diagram and the vascular topological feature vector, learning the mapping relation between different focus combination modes and the international clinical grading standard through an attention mechanism and generating a diabetic retinopathy grading result and a confidence score; The self-adaptive feedback optimization module is connected with the hierarchical decision module and the multi-focus cooperative detection module and is used for dynamically adjusting the detection threshold parameters of the multi-focus cooperative detection module according to the confidence score, and when the confidence score is lower than a preset threshold, focus detection sensitivity is enhanced to form a closed-loop feedback adjustment mechanism.
  2. 2. The diabetic retinopathy stratification diagnostic system of claim 1, wherein said multi-focal cooperative detection module comprises: The shared feature encoder is used for carrying out multi-scale feature extraction on the fundus image; A focus detection sub-network connected with the shared feature encoder and used for respectively detecting the positions and the categories of five focuses based on the multi-scale features; and the space association degree calculation unit is connected with the focus detection sub-network and is used for calculating the space proximity degree between different focus types and generating a focus collaborative perception index.
  3. 3. The hierarchical diagnosis system of diabetic retinopathy according to claim 2, wherein the spatial correlation calculation unit calculates euclidean distance between any two lesions based on center coordinates of the lesions, counts co-occurrence frequency of the lesions within a preset distance threshold, and determines the lesion co-perception index according to a ratio of the co-occurrence frequency of the lesions to total number of the lesions.
  4. 4. The diabetic retinopathy hierarchical diagnosis system according to claim 1, wherein the vascular topology map construction module includes: the blood vessel segmentation unit is used for carrying out blood vessel segmentation on the fundus image and generating a blood vessel binary mask; The topology extraction unit is connected with the blood vessel segmentation unit and is used for skeletonizing the blood vessel binary mask, extracting a blood vessel central line and identifying blood vessel nodes on the blood vessel central line, wherein the blood vessel nodes comprise branch nodes and end points, and the blood vessel central line segments between adjacent blood vessel nodes are used as blood vessel segments; And the graph embedding unit is connected with the topology extraction unit and is used for mapping the vessel nodes and the vessel segments into a graph structure, and performing feature learning through a graph convolution network to generate the vessel topology feature vector.
  5. 5. The hierarchical diagnosis system of diabetic retinopathy according to claim 4, wherein the map embedding unit assigns a node feature vector to each blood vessel node, the node feature vector including a node degree, a local blood vessel width, and a blood vessel branch angle at the node, aggregates neighborhood node features by a multi-layer map convolution operation, and finally generates the blood vessel topology feature vector by global pooling.
  6. 6. The diabetic retinopathy hierarchical diagnosis system of claim 1 wherein the hierarchical decision module comprises: the feature fusion unit is used for splicing the focus collaborative perception index, the spatial distribution thermodynamic diagram and the vessel topology feature vector to form fusion features; The attention weighting unit is connected with the feature fusion unit and is used for calculating importance weights of different feature dimensions in the fusion features through a self-attention mechanism; and the grading classification unit is connected with the attention weighting unit and is used for generating the diabetic retinopathy grading result and the confidence score based on the weighted fusion characteristics.
  7. 7. The hierarchical diagnosis system of diabetic retinopathy according to claim 6, characterized in that the attention weighting unit calculates attention weights by dot product operation of query vector, key vector and value vector, weights different dimensions of the fusion feature, highlighting feature dimensions related to the hierarchical level.
  8. 8. The diabetic retinopathy hierarchical diagnosis system of claim 1 wherein the adaptive feedback optimization module compares the confidence score with a preset high confidence threshold and low confidence threshold, decreases the detection sensitivity of the multi-modality cooperative detection module if the confidence score is above the high confidence threshold, and increases the detection sensitivity of the multi-modality cooperative detection module if the confidence score is below the low confidence threshold.
  9. 9. The diabetic retinopathy stratification diagnostic system of claim 1, further comprising: the image quality evaluation module is used for evaluating the quality of the fundus image, detecting the pupil size, the definition of the refractive medium and the image contrast, and outputting the re-shooting prompt information when the image quality does not meet the preset standard.
  10. 10. A method for the hierarchical diagnosis of diabetic retinopathy using the hierarchical diagnosis system of diabetic retinopathy according to any one of claims 1 to 9, characterized by comprising the steps of: based on the shared feature encoder, feature extraction is carried out on the eye bottom image, five focuses of microaneurysms, hard exudation, cotton-wool spots, bleeding foci and new blood vessels are identified, a spatial distribution thermodynamic diagram and a counting statistic of each focus are output, and focus collaborative perception indexes are calculated based on spatial adjacency and co-occurrence modes among the focuses; Performing blood vessel segmentation on the fundus image, extracting a blood vessel center line and branch nodes, constructing a blood vessel topological graph, taking the blood vessel nodes as graph nodes, taking the blood vessel segments as graph edges, performing embedded learning on the blood vessel topological graph through a graph neural network, and generating a blood vessel topological feature vector; Carrying out multi-scale feature fusion on the focus collaborative perception index, the spatial distribution thermodynamic diagram and the vessel topological feature vector, and learning the mapping relation between different focus combination modes and the international clinical grading standard through an attention mechanism to generate a diabetic retinopathy grading result and a confidence score; and dynamically adjusting the focus detection threshold parameter according to the confidence score, and enhancing focus detection sensitivity when the confidence score is lower than a preset threshold value to form a closed loop feedback adjustment mechanism.

Description

Grading diagnosis system and method for diabetic retinopathy Technical Field The invention relates to the technical field of medical image processing and computer vision, in particular to a diabetic retinopathy grading diagnosis system and method. Background Diabetic retinopathy is one of the most common microvascular complications of diabetes and is the leading cause of blindness in people of working age. According to international clinical diabetic retinopathy grading criteria, diabetic retinopathy can be classified into five grades, no obvious lesions, mild non-proliferative phase, moderate non-proliferative phase, severe non-proliferative phase and proliferative phase. Early detection and accurate classification are of great significance for timely intervention and delay of disease progression. At present, the basic medical institutions generally lack of professional ophthalmologists, and the problems of large workload, strong subjectivity, high missed diagnosis and misdiagnosis rate and the like exist in the screening of diabetic retinopathy by means of manual reading. In recent years, deep learning technology has made remarkable progress in the field of medical image analysis, and a new solution is provided for realizing automatic diagnosis of diabetic retinopathy. The existing automatic diagnosis method for diabetic retinopathy mainly has the following defects: Chinese patent application number 202510208120.9 discloses a lesion classification method employing a multi-branch network architecture, including a feature extractor, a multi-branch classifier, and a visual transducer. According to the method, characteristics of different scales are processed through a multi-branch structure, and the weight of each branch is dynamically adjusted by utilizing a vision converter so as to solve the problem of sample imbalance. However, the method has the following defects that firstly, only global feature learning at an image level is concerned, the cooperative relationship between local spatial distribution information of focuses and the focuses is not fully utilized, secondly, important influences of fundus blood vessel topological structures on diabetic retinopathy grading are ignored, blood vessel morphology and topology change are key indexes for judging the severity of the lesions, thirdly, an effective closed loop feedback mechanism is lacked, detection parameters cannot be dynamically adjusted according to the credibility of a diagnosis result, and therefore the diagnosis accuracy of a low-confidence sample is low. In summary, the prior art has the problems of insufficient utilization of lesion space associated information, insufficient mining of vascular topological features, insufficient adaptive optimization capability and the like in the grading diagnosis of diabetic retinopathy, and a grading diagnosis system and method for diabetic retinopathy, which can fully integrate lesion cooperative information and vascular topological features and have closed loop feedback optimization capability, are needed. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide a grading diagnosis system and a grading diagnosis method for diabetic retinopathy, which are characterized in that by constructing a deep coupling closed loop cooperative system for cooperative detection of multiple lesions, analysis of a vessel topological graph and grading decision, the spatial distribution characteristics and vessel topological structure information of the lesions are fully utilized, and detection parameters are dynamically adjusted through a self-adaptive feedback optimization mechanism, so that the accuracy and reliability of grading of the diabetic retinopathy are remarkably improved, and the actual requirements of fundus screening of basic medical institutions are met. In order to achieve the above purpose, the invention adopts the following technical scheme: The grading diagnosis system for the diabetic retinopathy comprises a multi-lesion cooperative detection module, a vascular topological graph construction module, a grading decision module and an adaptive feedback optimization module. The multi-focus collaborative detection module performs feature extraction on the eye bottom image based on the shared feature encoder, identifies five typical focuses, and outputs a spatial distribution thermodynamic diagram and focus collaborative perception indexes. The blood vessel topological graph construction module performs blood vessel segmentation and topology extraction on the fundus image, and learns the blood vessel topological characteristics through a graph neural network. The hierarchical decision module fuses the lesion features and the vessel topological features and generates a hierarchical result and a confidence score through an attention mechanism. The adaptive feedback optimization module dynamically adjusts the focus detection threshold according to the confidence score to form