Dynamic Graph Clustering Learning for Unsupervised Diabetic Retinopathy Classification

Abstract

Diabetic retinopathy (DR) is a common complication of diabetes, which can lead to vision loss. Early diagnosis is crucial to prevent the progression of DR. In recent years, deep learning approaches have shown promising results in the development of an intelligent and efficient system for DR classification. However, one major drawback is the need for expert-annotated datasets, which are both time-consuming and costly. To address these challenges, this paper proposes a novel dynamic graph clustering learning (DGCL) method for unsupervised classification of DR, which innovatively deploys the Euclidean and topological features from fundus images for dynamic clustering. Firstly, a multi-structural feature fusion (MFF) module extracts features from the structure of the fundus image and captures topological relationships among multiple samples, generating a fused representation. Secondly, another consistency smoothing clustering (CSC) module combines network updates and deep clustering to ensure stability and smooth performance improvement during model convergence, optimizing the clustering process by iteratively updating the network and refining the clustering results. Lastly, dynamic memory storage is utilized to track and store important information from previous iterations, enhancing the training stability and convergence. During validation, the experimental results with public datasets demonstrated the superiority of our proposed DGCL network.

Diabetic Retinopathy Lesion Segmentation Method Based on Multi-Scale Attention and Lesion Perception

“…Considering that the contrast between lesion areas and normal tissues in retinal fundus images is often imperceptible, it is difficult to distinguish between them with the naked eye and automated DR diagnosis based on deep learning [30]. Therefore, many methods [31,32] employ various techniques to highlight the lesion areas in the images before performing DR diagnosis, assisting the network to better process the images. Following the above-mentioned methods, this paper applies the following preprocessing steps to the DDR dataset, DIARETDB1 dataset, and IDRiD dataset: First, the black borders are cropped to reduce computation time and suppress irrelevant noise information.…”