CN-122023307-A - Self-supervision diabetic retinopathy grading method based on distraction and graph convolution

CN122023307ACN 122023307 ACN122023307 ACN 122023307ACN-122023307-A

Abstract

The invention belongs to the field of medical image processing and computer vision, and particularly relates to a self-supervision diabetic retinopathy grading method based on distraction and graph convolution. The method comprises the steps of introducing a self-supervision mechanism, utilizing effective information of a small quantity of marked samples to mine inherent identification features of unmarked samples, generating pseudo labels by utilizing high-confidence samples predicted by a model, and combining the self-supervision learning optimization model with feature enhancement consistency. The method has the advantages that a plurality of experimental results on APTOS2019 public data sets are better than other comparison methods under the same marked data, and the method is proved to have stronger competitiveness in semi-supervised DR grading.

Inventors

ZHANG CHENRUI
KANG JIANGUO
LI HUI
LIU YANG
WANG LIHUA
MENG DEJIA
WANG YONGJIE

Assignees

吕梁学院

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (4)

1. A self-supervision diabetic retinopathy grading method based on distraction and graph convolution is characterized by providing a semi-supervision learning model which comprises a feature extractor CNN, a graph rolling network GCN and a classifier F, mining association information among unlabeled samples through the graph rolling network GCN with the help of a limited number of labeling samples to obtain fine granularity features with more discrimination and achieve high efficiency of DR classification, designing a distraction module DM aiming at deploying the attention of the network to the focus area of fundus images in a targeted manner so as to extract fine discrimination features in the samples more effectively, and introducing an improved self-supervision mechanism SM for exploring intrinsic information in unlabeled samples according to the enhanced consistency similarity of unlabeled samples so as to improve the robustness of the model.
2. The self-supervised diabetic retinopathy stratification method based on distraction and graph convolution of claim 1 comprising a training phase and a testing phase, wherein a dual-phase semi-supervised training strategy is employed in the training phase, the first phase of training the distraction module DM using labeled samples and the second phase of combining labels with high confidence pseudo-label samples for consistency enhancement training in the self-supervision mechanism SM.
3. The self-supervised diabetic retinopathy stratification method based on distraction and graph convolution of claim 2 wherein the first phase of the training phase comprises the steps of: data preprocessing, namely, the input fundus image set is expressed as: Wherein: Is a labeled sample image; Is a label-free sample image; weak enhancement of each image And strength enhancement Two enhancement processes, the enhanced sample image is expressed as: , , ; The characteristic extraction network comprises the following extracted characteristics: , , ; In order to be a feature extractor of the present invention, A graph rolling network; Distraction module DM training: step 1-generating an attention mask from the feature map The maximum value channel is selected as an initial mask: , wherein, The attention mask is shown as such, Representing taking a maximum along the channel dimension, upsample representing an upsampling operation; Step 2, thresholding to generate a binary mask, namely setting a threshold value , , wherein, Representing a binary mask image; step 3, masking application, namely covering an original image by using a binary masking image, and highlighting a focus area: , wherein, Representing element-by-element multiplication; Representing the masked image; Step 4, feature re-extraction, namely inputting the masked image into a feature extractor CNN and a graph rolling network GCN again: Wherein, the method comprises the steps of, Representing the re-extracted features; step 5, loss function, using cross entropy loss supervision with label sample: , wherein, In order to be a classifier of the class, Representing a sample image with labels A corresponding tag; The distraction module DM is used for training the feature extractor CNN, the graph rolling network GCN and the classifier F and updating parameters of the feature extractor CNN, the graph rolling network GCN and the classifier F; Self-supervision mechanism SM: Step1, pseudo tag generation, namely predicting unlabeled samples: , , wherein, 、 Vectors representing predictions after weak and strong enhancement of unlabeled samples, respectively, with a selection confidence above a threshold Is used for the vector of (a), Wherein ; Step 2, enhancing consistency loss, namely constraining the feature consistency of the same sample under different enhancements: ; and 3, total loss function, namely performing end-to-end training by integrating various losses: wherein 、、 A balance coefficient indicating each loss; The self-supervision mechanism SM uses the high confidence samples to enhance the consistency self-supervision loss, and updates the parameters of the feature extractor CNN, the graph rolling network GCN and the classifier F to improve the DR classification performance.
4. A self-supervising diabetic retinopathy grading method based on distraction and graph convolution according to claim 2 or 3 wherein the trained model predicts lesion grading directly on the input data during the testing phase.

Description

Self-supervision diabetic retinopathy grading method based on distraction and graph convolution Technical Field The invention belongs to the field of medical image processing and computer vision, and particularly relates to a self-supervision diabetic retinopathy grading method based on distraction and graph convolution. Background Diabetic retinopathy (Diabetic Retinopathy, DR) is a microvascular complication caused by diabetes, DR can cause damage to the ocular nerves, resulting in impaired and even blindness of the patient's vision, and therefore early diagnosis and discovery of the disease is critical to the patient. The severity of DR is generally dependent on a combination of lesion characteristics such as microaneurysms, hemorrhages, hard and soft exudates, and according to international standards, the severity of DR can be categorized into five different classes of no, mild, moderate, severe and proliferative. In addition, there are two categories, DR-free and DR, according to some accepted standards. In clinical practice, retinal disease diagnosis faces significant challenges, and the time-consuming and lengthy disease grading diagnosis and treatment procedures are caused by the existence of multiple layers of image data in the retinal image and the subtle morphological differences between the grades. The long-time and high-intensity image interpretation work easily causes visual fatigue and cognitive load to clinical specialists, causes adverse events such as misdiagnosis, missed diagnosis and the like, and seriously influences the diagnosis accuracy and the effectiveness of treatment decisions. Such problems not only restrict clinical work efficiency, but also constitute a potential threat to timely intervention and prognosis management of patients. Therefore, the development of an automated DR diagnostic method is of great practical significance. The existing DR automatic grading method based on deep learning mostly adopts full-supervised learning, a large number of fundus images are required to be accurately marked, marking work is time-consuming and labor-consuming, cost is high, and popularization and application of the model in clinic are limited. Disclosure of Invention Aiming at the problem that a large amount of fundus image data labeling work is required by the existing DR automatic grading method, the invention provides a self-supervision diabetic retinopathy grading method based on distraction and graph convolution, which aims at reducing the labeling workload of ophthalmologists, and the method enables a model frame to focus at multiple angles and excavate more efficient fine granularity characteristics by introducing a distraction module DM; in addition, a self-supervision mechanism SM is introduced, the clustering effect of the model framework is improved, and the dependence of the model framework on the labeling samples is reduced. The invention adopts the following technical scheme that a self-supervision diabetic retinopathy grading method based on distraction and graph convolution provides a semi-supervision learning model, the model comprises a feature extractor CNN, a graph rolling network GCN and a classifier F, under the help of a limited number of marked samples, the graph rolling network GCN is used for mining the association information among unmarked samples to obtain fine granularity characteristics with better discrimination and realize the high efficiency of DR classification, a distraction module DM is designed for deploying the attention of the network to the focus area of a fundus image in a targeted manner so as to extract fine discrimination characteristics in the samples more effectively, an improved self-supervision mechanism SM is introduced, the intrinsic information in the unmarked samples is explored according to the enhanced consistency similarity of the unmarked samples, and the robustness of the model is improved. The self-supervision diabetic retinopathy grading method based on distraction and graph convolution comprises a training stage and a testing stage, wherein a double-stage semi-supervision training strategy is adopted in the training stage, namely a first stage of training a distraction module DM by using a labeled sample and a second stage of carrying out consistency enhancement training in a self-supervision mechanism SM by combining a label and a high-confidence pseudo-label sample. The self-monitoring diabetic retinopathy grading method based on distraction and graph convolution comprises the following steps: data preprocessing, namely, the input fundus image set is expressed as: Wherein: Is a labeled sample image; Is a label-free sample image; weak enhancement of each image And strength enhancementTwo enhancement processes, the enhanced sample image is expressed as:,,; The characteristic extraction network comprises the following extracted characteristics: ,,; In order to be a feature extractor of the present invention, A graph rolling