Search

CN-122023950-A - Automatic screening system for fundus color illumination sugar net lesions based on image identification

CN122023950ACN 122023950 ACN122023950 ACN 122023950ACN-122023950-A

Abstract

The invention relates to the technical field of medical image processing and artificial intelligence, and particularly discloses an automatic screening system for fundus color illumination sugar net lesions based on image identification. The system comprises a multi-mode feature decoupling extraction module, a specific lesion feature strengthening module, a dynamic decision boundary generation module and a hierarchical cascade classification and confidence assessment module. According to the invention, by constructing the multi-mode feature decoupling extraction module, the global structure information and the local focus information which are easy to be confused are physically isolated and processed in parallel in the feature extraction stage from the technical source, so that the mutual interference of different pathological semantic features in the early encoding stage is restrained, and a solid foundation is laid for the subsequent generation of high-specificity features.

Inventors

  • YANG HUAJING
  • NI LIANG

Assignees

  • 华中科技大学同济医学院附属同济医院

Dates

Publication Date
20260512
Application Date
20260413

Claims (10)

  1. 1. Automatic screening system of eyeground color photograph sugar net pathological change based on image discrimination, its characterized in that includes: The multi-mode feature decoupling extraction module is used for receiving an input fundus color image and synchronously executing at least two parallel feature extraction paths to generate feature vectors with different pathological semantic orientations; the specific lesion feature strengthening module is used for carrying out cross-path contrast analysis and self-adaptive weighted fusion on the feature vector generated by the multi-mode feature decoupling extraction module so as to generate a comprehensive feature representation with high sugar network specificity; the dynamic decision boundary generation module is used for calculating and generating a classification decision boundary matched with the image feature distribution in real time based on the comprehensive feature representation of the fundus color image which is input currently; and the hierarchical cascade classification and confidence evaluation module is used for carrying out multi-stage classification judgment on the comprehensive characteristic representation according to the decision boundary provided by the dynamic decision boundary generation module and synchronously outputting a lesion level classification result and a corresponding confidence score.
  2. 2. The automatic screening system for fundus illumination sugar network lesions based on image discrimination according to claim 1, wherein said multi-modal feature decoupling extraction module comprises a global structural feature extraction path and a local lesion feature extraction path; The global structural feature extraction path is provided with a first depth convolution neural network, the first depth convolution neural network takes a whole fundus color image as input, and the network architecture of the first depth convolution neural network is optimized through pre-training so as to capture the overall form, relative position relation and macroscopic texture distribution information of the optic disc, the macula lutea and the main vessel arch in a key way; The local focus characteristic extraction path is configured with a second deep convolution neural network and a focus candidate region suggestion network, the local focus characteristic extraction path firstly automatically generates a plurality of candidate focus regions in an input image through the focus candidate region suggestion network, and then the candidate regions are subjected to high-resolution clipping and characteristic coding by the second deep convolution neural network so as to specially extract micro-morphology and texture details of micro-aneurysms, bleeding points and hard exudate local focuses.
  3. 3. The automatic screening system for fundus color illumination sugar net lesions based on image discrimination according to claim 2, wherein the workflow of the specific lesion characterization module is as follows: Receiving a global feature vector from the global structural feature extraction path and a local feature vector set from the local focus feature extraction path; Mapping each local feature vector to the same dimension as the global feature vector through a linear transformation layer by a built-in cross-path feature comparison unit, then calculating cosine similarity between the global feature vector and the local feature vector aligned with each dimension, and dividing the local feature vector into a feature subset with high association degree with the global structure and a feature subset with low association degree according to a preset similarity threshold; different fusion weights are respectively applied to the two feature subsets through a built-in self-adaptive feature fusion unit, higher fusion weights are given to the local feature subsets with low association degree with the global structure, and lower fusion weights are given to the local feature subsets with high association degree with the global structure; And aggregating the weighted local feature subsets, and splicing the local feature subsets with the global feature vectors to generate the comprehensive feature representation.
  4. 4. The automatic screening system for fundus color illumination sugar network lesions based on image discrimination according to claim 3, wherein the core of said dynamic decision boundary generation module is a boundary calculation network; The boundary computing network takes the comprehensive characteristic representation output by the specific lesion characteristic strengthening module as input; the boundary computing network carries out nonlinear transformation through 3 full-connection layers and finally outputs decision boundary parameter vectors which define normal vectors and bias items of hyperplanes in a high-dimensional feature space; The hyperplane is used as a real-time classification decision boundary for the current input sample, and the position and the direction of the hyperplane are dynamically determined by the comprehensive characteristics of the input sample.
  5. 5. The automatic screening system for fundus illumination sugar network lesions based on image discrimination according to claim 4, wherein said hierarchical cascade classification and confidence assessment module comprises a cascade classifier stack and confidence assessor; The cascade classifier stack is formed by sequentially connecting a plurality of shallow classifiers, and the 1 st class classifier receives the comprehensive characteristic representation and performs preliminary classification judgment according to a dynamic decision boundary, namely positive or negative sugar network lesions; If the class 1 classifier judges positive, the comprehensive characteristic representation and the class 1 judgment result are transmitted to a class 2 classifier, and the class 2 classifier is responsible for classifying the lesion severity degree in a positive sample and outputting a classification result according to an international clinical classification standard; The confidence level evaluator works in parallel with each class of classifier, receives the last hidden layer activation value of the current class of classifier before outputting the final class, calculates the minimum Euclidean distance between the hidden layer activation value vector and each prototype vector, and maps the minimum Euclidean distance to a value between 0 and 1 through a preset monotonically decreasing function to be used as a confidence level score of the current classification result.
  6. 6. The automatic fundus color illumination sugar network lesion screening system based on image discrimination according to claim 5, further comprising a model persistence optimization interface; the model continuous optimization interface is used for receiving screening result feedback confirmed by the professional doctor after the system is deployed; The model continuous optimization interface packages feedback data comprising an original fundus image, a comprehensive feature representation generated by a system, a lesion label corrected by a doctor and a confidence score output by the system as a training sample pair; The training sample pairs are stored in an incremental learning buffer pool, and when the number of samples in the buffer pool reaches a preset threshold, incremental fine tuning training of the multi-mode characteristic decoupling extraction module, the specific lesion characteristic strengthening module and the layering cascade classification and confidence assessment module is triggered.
  7. 7. The automatic screening system for fundus color illumination sugar network lesions based on image discrimination according to claim 6, wherein the operation mechanism of the lesion candidate area suggestion network is as follows: generating dense anchor point grids on an input image, and predicting the existence probability score of a focus and the position adjustment quantity of a boundary frame for each anchor point; a non-maximum suppression algorithm is adopted, and a plurality of candidate areas which are ranked at the front are screened out according to the lesion existence probability score; And precisely cutting out a corresponding image block from the original image through a space transformation layer of the focus candidate region according to the coordinate information of the candidate region, and processing the image block by a subsequent second depth convolution neural network.
  8. 8. The automatic fundus color illumination sugar network lesion screening system based on image discrimination according to claim 7, wherein said system is run in compliance with a preset confidence-workflow linkage protocol; the confidence level-workflow linkage protocol provides that when the confidence level score of the highest-level lesion classification result output by the hierarchical cascade classification and confidence level assessment module is smaller than a first preset threshold value, the system automatically marks the case as needing to be manually rechecked, and pushes the images and all the intermediate feature visualization results of the case to a manual rechecked queue; When the confidence score is smaller than a second preset threshold value, the system automatically activates an internal feature re-extraction process in addition to marking the required manual re-verification, and in the process, the specific lesion feature enhancement module regenerates the comprehensive feature representation by adopting a group of alternative and more conservative feature fusion weight coefficients and classifies and evaluates the comprehensive feature representation again.
  9. 9. The automatic screening system for fundus illumination sugar network lesions based on image discrimination according to claim 8, wherein the lesion candidate region suggestion network in the local lesion feature extraction path adopts a multi-scale feature pyramid mechanism; the multi-scale feature pyramid mechanism utilizes the output of an intermediate convolution layer of the first deep convolution neural network to construct a top-down feature pyramid; the focus candidate region suggestion network deploys anchor point grids on each layer of the feature pyramid independently, and the anchor point size of each layer is scaled according to the receptive field of each layer of feature map; after independent non-maximum value inhibition, the region proposal results of each layer are combined and secondarily screened on a unified scale to generate candidate regions.
  10. 10. The automatic screening system for fundus color illumination sugar net lesions based on image discrimination according to claim 9, wherein the boundary computing network embeds a channel attention module between the second fully connected layer and the third fully connected layer; the channel attention module carries out global average pooling on the output of the second full-connection layer to generate a channel statistical vector; The channel attention module carries out nonlinear transformation on the channel statistical vector through two 1X 1 convolution layers to generate channel weight; The channel attention module multiplies the channel weight and the original feature channel by channel to obtain a re-weighted feature representation for processing by a third full connection layer.

Description

Automatic screening system for fundus color illumination sugar net lesions based on image identification Technical Field The invention belongs to the technical field of medical image processing and artificial intelligence, and particularly relates to an automatic screening system for fundus color illumination sugar net lesions based on image identification. Background The artificial intelligence technology plays an increasingly important role in the field of medical image analysis, assists doctors in diagnosing diseases through methods such as deep learning and the like, and remarkably improves diagnosis and treatment efficiency and accuracy. The automatic screening of diabetic retinopathy based on fundus color photography is a key application direction of artificial intelligence auxiliary diagnosis, and aims to realize early and rapid identification of the diabetic retinopathy by analyzing fundus images. In the prior art, a convolutional neural network and other models are generally adopted to perform feature extraction and classification on single fundus color illumination. However, certain pathological characterizations of sugar network lesions, such as microaneurysms, hemorrhagic spots, etc., have visual similarities in images with fundus changes caused by other systemic diseases such as hypertensive retinopathy, which results in easy confusion of models in feature learning, and difficult extraction of highly specific discriminating features. The characteristic confusion problem directly affects the performance of the classification model, so that the existing automatic screening system has higher misjudgment risk when facing complex and diverse clinical actual images, and the reliability and generalization capability of the screening result are urgently improved. Disclosure of Invention The invention aims to provide an automatic screening system for fundus color illumination sugar net lesions based on image discrimination, which is used for solving the problems of model feature confusion, screening specificity and insufficient generalization capability caused by visual similarity of the sugar net lesion features and other fundus lesion features in the prior art. The invention provides an automatic screening system for fundus color illumination sugar net lesions based on image identification, which comprises: The multi-mode feature decoupling extraction module is used for receiving an input fundus color image and synchronously executing at least two parallel feature extraction paths to generate feature vectors with different pathological semantic orientations; the specific lesion feature strengthening module is used for carrying out cross-path contrast analysis and self-adaptive weighted fusion on the feature vector generated by the multi-mode feature decoupling extraction module so as to generate a comprehensive feature representation with high sugar network specificity; the dynamic decision boundary generation module is used for calculating and generating a classification decision boundary matched with the image feature distribution in real time based on the comprehensive feature representation of the fundus color image which is input currently; and the hierarchical cascade classification and confidence evaluation module is used for carrying out multi-stage classification judgment on the comprehensive characteristic representation according to the decision boundary provided by the dynamic decision boundary generation module and synchronously outputting a lesion level classification result and a corresponding confidence score. Preferably, the multi-modal feature decoupling extraction module includes a global structural feature extraction path and a local focus feature extraction path; The global structural feature extraction path is provided with a first depth convolution neural network, the first depth convolution neural network takes a whole fundus color image as input, and the network architecture of the first depth convolution neural network is optimized through pre-training so as to capture the overall form, relative position relation and macroscopic texture distribution information of the optic disc, the macula lutea and the main vessel arch in a key way; The local focus characteristic extraction path is configured with a second deep convolution neural network and a focus candidate region suggestion network, the local focus characteristic extraction path firstly automatically generates a plurality of candidate focus regions in an input image through the focus candidate region suggestion network, and then the candidate regions are subjected to high-resolution clipping and characteristic coding by the second deep convolution neural network so as to specially extract micro-morphology and texture details of micro-aneurysms, bleeding points and hard exudate local focuses. Preferably, the workflow of the specific lesion characterization enhancement module is as follows: Receiving a global feature vecto