Dynamic Graph Clustering Learning for Unsupervised Diabetic Retinopathy Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Early diagnosis of DR is critical for preventing irreversible vision impairment.
• Current deep learning methods for DR classification often require large, expert-annotated datasets, which are expensive and time-consuming to create.

Purpose of the Study:

• To develop a novel unsupervised deep learning method for diabetic retinopathy classification.
• To address the limitations of supervised learning approaches, specifically the reliance on expert-annotated data.
• To improve the efficiency and accessibility of DR screening systems.

Main Methods:

• Proposes a dynamic graph clustering learning (DGCL) method for unsupervised DR classification.
• Employs a multi-structural feature fusion (MFF) module to extract Euclidean and topological features from fundus images.
• Utilizes a consistency smoothing clustering (CSC) module and dynamic memory storage for stable and improved clustering performance.

Main Results:

• The proposed DGCL method demonstrated superior performance in unsupervised DR classification on public datasets.
• The method effectively extracts and fuses structural and topological features for enhanced classification.
• Achieved stable model convergence and improved clustering accuracy without expert annotations.

Conclusions:

• The DGCL method offers a promising unsupervised approach for DR classification, reducing reliance on manual annotation.
• This technique has the potential to significantly improve early DR detection and management, especially in resource-limited settings.
• The innovative feature fusion and clustering stability mechanisms pave the way for more efficient AI-driven ophthalmic diagnostics.